Datasets:

jamescalam

/

llama-2-arxiv-papers-chunked

like 17

1512.02595

50

Mandarin and accented Mandarin. 5.1 Dataset Construction Some of the internal English (3,600 hours) and Mandarin (1,400 hours) datasets were created from raw data captured as long audio clips with noisy transcriptions. The length of these clips ranged from several minutes to more than hour, making it impractical to unroll them in time in the RNN during training. To solve this problem, we developed an alignment, segmentation and filtering pipeline that can generate a training set with shorter utterances and few erroneous transcriptions. The first step in the pipeline is to use an existing bidirectional RNN model trained with CTC to align the transcription to the frames of audio. For a given audio-transcript pair, (x,y), we find the alignment that maximizes `= arg max `2Align (x,y)TY tpctc(`tjx;). (13) This is essentially a Viterbi alignment found using a RNN model trained with CTC. Since Equation 9 integrates over the alignment, the CTC loss function is never explicitly asked to produce an accurate alignment. In principle, CTC could choose to emit all the characters of the transcription after some 15

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

51

alignment. In principle, CTC could choose to emit all the characters of the transcription after some 15 fixed delay and this can happen with unidirectional RNNs [54]. However, we found that CTC produces an accurate alignment when trained with a bidirectional RNN. Following the alignment is a segmentation step that splices the audio and the corresponding aligned transcription whenever it encounters a long series of consecutive blank labels occurs, since this usually denotes a stretch of silence. By tuning the number of consecutive blank s, we can tune the length of the utterances generated. For the English speech data, we also require a space token to be within the stretch of blank s in order to segment only on word boundaries. We tune the segmentation to generate utterances that are on average 7 seconds long. The final step in the pipeline removes erroneous examples that arise from a failed alignment. We crowd source the ground truth transcriptions for several thousand examples. The word level edit distance between the ground truth and the aligned transcription is used to produce a good orbad label. The threshold for the word level edit distance is chosen such that the resulting WER of the good portion of the development set is less than 5%. We then train a linear classifier to accurately

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

52

good portion of the development set is less than 5%. We then train a linear classifier to accurately predict bad examples given the input features generated from the speech recognizer. We find the following features useful: the raw CTC cost, the CTC cost normalized by the sequence length, the CTC cost normalized by the transcript length, the ratio of the sequence length to the transcript length, the number of words in the transcription and the number of characters in the transcription. For the English dataset, we find that the filtering pipeline reduces the WER from 17% to 5% while retaining more than 50% of the examples. 5.2 Data Augmentation We augment our training data by adding noise to increase the effective size of our training data and to improve our robustness to noisy speech [26]. Although the training data contains some intrinsic noise, we can increase the quantity and variety of noise through augmentation. Too much noise augmentation tends to make optimization difficult and can lead to worse results, and too little noise augmentation makes the system less robust to low signal-to-noise speech. We find that a good balance is to add noise to 40% of the utterances that are chosen at random. The noise source consists

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

53

balance is to add noise to 40% of the utterances that are chosen at random. The noise source consists of several thousand hours of randomly selected audio clips combined to produce hundreds of hours of noise. 5.3 Scaling Data Our English and Mandarin corpora are substantially larger than those commonly reported in speech recognition literature. In Table 10, we show the effect of increasing the amount of labeled training data on WER. This is done by randomly sampling the full dataset before training. For each dataset, the model was trained for up to 20 epochs though usually early-stopped based on the error on a held out development set. We note that the WER decreases with a power law for both the regular and noisy development sets. The WER decreases by 40% relative for each factor of 10 increase in training set size. We also observe a consistent gap in WER ( 60% relative) between the regular and noisy datasets, implying that more data benefits both cases equally. This implies that a speech system will continue to improve with more labeled training data. We hypothesize that equally as important as increasing raw number of hours is increasing the number of speech contexts that are captured in the dataset. A context can be any property that makes speech unique including different speakers, background noise, environment, and microphone hardware.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

54

of speech contexts that are captured in the dataset. A context can be any property that makes speech unique including different speakers, background noise, environment, and microphone hardware. While we do not have the labels needed to validate this claim, we suspect that measuring WER as a function of speakers in the dataset would lead to much larger relative gains than simple random sampling. 6 Results To better assess the real-world applicability of our speech system, we evaluate on a wide range of test sets. We use several publicly available benchmarks and several test sets collected internally. Together these test sets represent a wide range of challenging speech environments including low signal-to-noise ratios (noisy and far-field), accented, read, spontaneous and conversational speech. 16 Fraction of Data Hours Regular Dev Noisy Dev 1% 120 29.23 50.97 10% 1200 13.80 22.99 20% 2400 11.65 20.41 50% 6000 9.51 15.90 100% 12000 8.46 13.59 Table 10: Comparison of English WER for Regular and Noisy development sets on increasing training dataset size. The architecture is a 9-layer model with 2 layers of 2D-invariant convolution and 7 recurrent layers with 68M parameters.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

55

size. The architecture is a 9-layer model with 2 layers of 2D-invariant convolution and 7 recurrent layers with 68M parameters. All models are trained for 20 epochs on either the full English dataset, described in Table 9, or the full Mandarin dataset described in Section 5. We use stochastic gradient descent with Nesterov momentum [61] along with a minibatch of 512 utterances. If the norm of the gradient exceeds a threshold of 400, it is rescaled to 400 [47]. The model which performs the best on a held-out development set during training is chosen for evaluation. The learning rate is chosen from [1 104, 6104]to yield fastest convergence and annealed by a constant factor of 1.2 after each epoch. We use a momentum of 0.99 for all models. The language models used are those described in Section 3.8. The decoding parameters from Equation 12 are tuned on a held-out development set. We use a beam size of 500 for the English decoder and a beam size of 200 for the Mandarin decoder. 6.1 English The best DS2 model has 11 layers with 3 layers of 2D convolution, 7 bidirectional recurrent layers,

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

56

6.1 English The best DS2 model has 11 layers with 3 layers of 2D convolution, 7 bidirectional recurrent layers, a fully-connected output layer along with Batch Normalization. The first layer outputs to bigrams with a temporal stride of 3. By comparison the DS1 model has 5 layers with a single bidirectional recurrent layer and it outputs to unigrams with a temporal stride of 2 in the first layer. We report results on several test sets for both the DS2 and DS1 model. We do not tune or adapt either model to any of the speech conditions in the test sets. Language model decoding parameters are set once on a held-out development set. To put the performance of our system in context, we benchmark most of our results against human workers, since speech recognition is an audio perception and language understanding problem that humans excel at. We obtain a measure of human level performance by paying workers from Amazon Mechanical Turk to hand-transcribe all of our test sets. Two workers transcribe the same audio clip, that is typically about 5 seconds long, and we use the better of the two transcriptions for the final WER calculation. They are free to listen to the audio clip as many times as they like. These workers

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

57

WER calculation. They are free to listen to the audio clip as many times as they like. These workers are mostly based in the United States, and on average spend about 27 seconds per transcription. The hand-transcribed results are compared to the existing ground truth to produce a WER. While the existing ground truth transcriptions do have some label error, this is rarely more than 1%. This implies that disagreement between the ground truth transcripts and the human level transcripts is a good heuristic for human level performance. 6.1.1 Model Size Our English speech training set is substantially larger than the size of commonly used speech datasets. Furthermore, the data is augmented with noise synthesis. To get the best generalization error, we expect that the model size must increase to fully exploit the patterns in the data. In Section 3.2 we explored the effect of model depth while fixing the number of parameters. In contrast, here we show the effect of varying model size on the performance of the speech system. We only vary the size of each layer, while keeping the depth and other architectural parameters constant. We evaluate the models on the same Regular and Noisy development sets that we use in Section 3.5. The models in Table 11 differ from those in Table 3 in that we increase the the stride to 3 and output

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

58

The models in Table 11 differ from those in Table 3 in that we increase the the stride to 3 and output to bigrams. Because we increase the model size to as many as 100 million parameters, we find that an increase in stride is necessary for fast computation and memory constraints. However, in this regime we note that the performance advantage of the GRU networks appears to diminish over the 17 Model size Model type Regular Dev Noisy Dev 18106GRU 10.59 21.38 38106GRU 9.06 17.07 70106GRU 8.54 15.98 70106RNN 8.44 15.09 100106GRU 7.78 14.17 100106RNN 7.73 13.06 Table 11: Comparing the effect of model size on the WER of the English speech system on both the regular and noisy development sets. We vary the number of hidden units in all but the convolutional layers. The GRU model has 3 layers of bidirectional GRUs with 1 layer of 2D-invariant convolution. The RNN model has 7 layers of bidirectional simple recurrence with 3 layers of 2D-invariant convolution. Both models output bigrams with a temporal stride of 3. All models contain approximately 35 million parameters and are trained with BatchNorm

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

59

temporal stride of 3. All models contain approximately 35 million parameters and are trained with BatchNorm and SortaGrad. Test set DS1 DS2 Baidu Test 24.01 13.59 Table 12: Comparison of DS1 and DS2 WER on an internal test set of 3,300 examples. The test set contains a wide variety of speech including accents, low signal-to-noise speech, spontaneous and conversational speech. simple RNN. In fact, for the 100 million parameter networks the simple RNN performs better than the GRU network and is faster to train despite the 2 extra layers of convolution. Table 11 shows that the performance of the system improves consistently up to 100 million parameters. All further English DS2 results are reported with this same 100 million parameter RNN model since it achieves the lowest generalization errors. Table 12 shows that the 100 million parameter RNN model (DS2) gives a 43.4% relative improvement over the 5-layer model with 1 recurrent layer (DS1) on an internal Baidu dataset of 3,300 utterances that contains a wide variety of speech including challenging accents, low signal-to-noise ratios from far-field or background noise, spontaneous and conversational speech. 6.1.2 Read Speech

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

60

ratios from far-field or background noise, spontaneous and conversational speech. 6.1.2 Read Speech Read speech with high signal-to-noise ratio is arguably the easiest large vocabulary for a continuous speech recognition task. We benchmark our system on two test sets from the Wall Street Journal (WSJ) corpus of read news articles. These are available in the LDC catalog as LDC94S13B and LDC93S6B. We also take advantage of the recently developed LibriSpeech corpus constructed using audio books from the LibriV ox project [46]. Table 13 shows that the DS2 system outperforms humans in 3 out of the 4 test sets and is competitive on the fourth. Given this result, we suspect that there is little room for a generic speech system to further improve on clean read speech without further domain adaptation. Read Speech Test set DS1 DS2 Human WSJ eval’92 4.94 3.60 5.03 WSJ eval’93 6.94 4.98 8.08 LibriSpeech test-clean 7.89 5.33 5.83 LibriSpeech test-other 21.74 13.25 12.69 Table 13: Comparison of WER for two speech systems and human level performance on read speech. 18 Accented Speech Test set DS1 DS2 Human

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

61

Table 13: Comparison of WER for two speech systems and human level performance on read speech. 18 Accented Speech Test set DS1 DS2 Human V oxForge American-Canadian 15.01 7.55 4.85 V oxForge Commonwealth 28.46 13.56 8.15 V oxForge European 31.20 17.55 12.76 V oxForge Indian 45.35 22.44 22.15 Table 14: Comparing WER of the DS1 system to the DS2 system on accented speech. Noisy Speech Test set DS1 DS2 Human CHiME eval clean 6.30 3.34 3.46 CHiME eval real 67.94 21.79 11.84 CHiME eval sim 80.27 45.05 31.33 Table 15: Comparison of DS1 and DS2 system on noisy speech. “CHiME eval clean” is a noise-free baseline. The “CHiME eval real” dataset is collected in real noisy environments and the “CHiME eval sim” dataset has similar noise synthetically added to clean speech. Note that we use only one of the six channels to test each utterance. 6.1.3 Accented Speech Our source for accented speech is the publicly available V oxForge ( http://www.voxforge.org )

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

62

utterance. 6.1.3 Accented Speech Our source for accented speech is the publicly available V oxForge ( http://www.voxforge.org ) dataset, which has clean speech read from speakers with many different accents. We group these accents into four categories. The American-Canadian and Indian groups are self-explanatory. The Commonwealth accent denotes speakers with British, Irish, South African, Australian and New Zealand accents. The European group contains speakers with accents from countries in Europe that do not have English as a first language. We construct a test set from the V oxForge data with 1024 examples from each accent group for a total of 4096 examples. Performance on these test sets is to some extent a measure of the breadth and quality of our training data. Table 14 shows that our performance improved on all the accents when we include more accented training data and use an architecture that can effectively train on that data. However human level performance is still notably better than that of DS2 for all but the Indian accent. 6.1.4 Noisy Speech We test our performance on noisy speech using the publicly available test sets from the recently completed third CHiME challenge [4]. This dataset has 1320 utterances from the WSJ test set

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

63

We test our performance on noisy speech using the publicly available test sets from the recently completed third CHiME challenge [4]. This dataset has 1320 utterances from the WSJ test set read in various noisy environments, including a bus, a cafe, a street and a pedestrian area. The CHiME set also includes 1320 utterances with simulated noise from the same environments as well as the control set containing the same utterances delivered by the same speakers in a noise-free environment. Differences between results on the control set and the noisy sets provide a measure of the network’s ability to handle a variety of real and synthetic noise conditions. The CHiME audio has 6 channels and using all of them can provide substantial performance improvements [69]. We use a single channel for all our results, since multi-channel audio is not pervasive on most devices. Table 15 shows that DS2 substantially improves upon DS1, however DS2 is worse than human level performance on noisy data. The relative gap between DS2 and human level performance is larger when the data comes from a real noisy environment instead of synthetically adding noise to clean speech. 6.2 Mandarin In Table 16 we compare several architectures trained on the Mandarin Chinese speech, on a development set of 2000 utterances as well as a test set of 1882 examples of noisy speech. This development

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

64

In Table 16 we compare several architectures trained on the Mandarin Chinese speech, on a development set of 2000 utterances as well as a test set of 1882 examples of noisy speech. This development set was also used to tune the decoding parameters We see that the deepest model with 2D-invariant 19 convolution and BatchNorm outperforms the shallow RNN by 48% relative, thus continuing the trend that we saw with the English system—multiple layers of bidirectional recurrence improves performance substantially. Architecture Dev Test 5-layer, 1 RNN 7.13 15.41 5-layer, 3 RNN 6.49 11.85 5-layer, 3 RNN + BatchNorm 6.22 9.39 9-layer, 7 RNN + BatchNorm + 2D conv 5.81 7.93 Table 16: Comparison of the improvements in DeepSpeech with architectural improvements. The development and test sets are Baidu internal corpora. All the models in the table have about 80 million parameters each We find that our best Mandarin Chinese speech system transcribes short voice-query like utterances better than a typical Mandarin Chinese speaker. To benchmark against humans we ran a test with 100 randomly selected utterances and had a group of 5 humans label all of them together. The group

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

65

better than a typical Mandarin Chinese speaker. To benchmark against humans we ran a test with 100 randomly selected utterances and had a group of 5 humans label all of them together. The group of humans had an error rate of 4.0% as compared to the speech systems performance of 3.7%. We also compared a single human transcriber to the speech system on 250 randomly selected utterances. In this case the speech system performs much better: 9.7% for the human compared to 5.7% for the speech model. 7 Deployment Real-world applications usually require a speech system to transcribe in real time or with relatively low latency. The system used in Section 6.1 is not well-designed for this task, for several reasons. First, since the RNN has several bidirectional layers, transcribing the first part of an utterance requires the entire utterance to be presented to the RNN. Second, since we use a wide beam when decoding with a language model, beam search can be expensive, particularly in Mandarin where the number of possible next characters is very large (around 6000). Third, as described in Section 3, we normalize power across an entire utterance, which again requires the entire utterance to be available in advance. We solve the power normalization problem by using some statistics from our training set to perform

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

66

in advance. We solve the power normalization problem by using some statistics from our training set to perform an adaptive normalization of speech inputs during online transcription. We can solve the other problems by modifying our network and decoding procedure to produce a model that performs almost as well while having much lower latency. We focus on our Mandarin system since some aspects of that system are more challenging to deploy (e.g. the large character set), but the same techniques could also be applied in English. In this section, latency refers to the computational latency of our speech system as measured from the end of an utterance until the transcription is produced. This latency does not include data transmission over the internet, and does not measure latency from the beginning of an utterance until the first transcription is produced. We focus on latency from end of utterance to transcription because it is important to applications using speech recognition. 7.1 Batch Dispatch In order to deploy our relatively large deep neural networks at low latency, we have paid special attention to efficiency during deployment. Most internet applications process requests individually as they arrive in the data center. This makes for a straightforward implementation where each request can be managed by one thread. However, processing requests individually is inefficient computationally, for two main reasons. Firstly, when processing requests individually, the processor must

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

67

can be managed by one thread. However, processing requests individually is inefficient computationally, for two main reasons. Firstly, when processing requests individually, the processor must load all the weights of the network for each request. This lowers the arithmetic intensity of the workload, and tends to make the computation memory bandwidth bound, as it is difficult to effectively use on-chip caches when requests are presented individually. Secondly, the amount of parallelism that can be exploited to classify one request is limited, making it difficult to exploit SIMD or multicore parallelism. RNNs are especially challenging to deploy because evaluating RNNs sample by 20 0 1 2 3 4 5 6 7 8 9 10 11 Batch size0.00.10.20.30.4Probability10 streams 20 streams 30 streamsFigure 5: Probability that a request is processed in a batch of given size sample relies on sequential matrix vector multiplications, which are bandwidth bound and difficult to parallelize. To overcome these issues, we built a batching scheduler called Batch Dispatch that assembles streams of data from user requests into batches before performing forward propagation on these batches. In this case, there is a tradeoff between increased batch size, and consequently improved

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

68

streams of data from user requests into batches before performing forward propagation on these batches. In this case, there is a tradeoff between increased batch size, and consequently improved efficiency, and increased latency. The more we buffer user requests to assemble a large batch, the longer users must wait for their results. This places constraints on the amount of batching we can perform. We use an eager batching scheme that processes each batch as soon as the previous batch is completed, regardless of how much work is ready by that point. This scheduling algorithm has proved to be the best at reducing end-user latency, despite the fact that it is less efficient computationally, since it does not attempt to maximize batch size. Figure 5 shows the probability that a request is processed in a batch of given size for our production system running on a single NVIDIA Quadro K1200 GPU, with 10-30 concurrent user requests. As expected, batching works best when the server is heavily loaded: as load increases, the distribution shifts to favor processing requests in larger batches. However, even with a light load of only 10 concurrent user requests, our system performs more than half the work in batches with at least 2 samples. 0 10 20 30 40 Number of concurrent streams050100latency (ms)50%ile 98%ile

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

69

samples. 0 10 20 30 40 Number of concurrent streams050100latency (ms)50%ile 98%ile Figure 6: Median and 98 percentile latencies as a function of server load We see in Figure 6, that our system achieves a median latency of 44 ms, and a 98 percentile latency of 70 ms when loaded with 10 concurrent streams. As the load increases on the server, the batching scheduler shifts work to more efficient batches, which keeps latency low. This shows that Batch Dispatch makes it possible to deploy these large models at high throughput and low latency. 21 1 2 3 4 5 6 7 8 9 10 Batch size0.00.10.20.30.40.5TeraFLOP/snervana baiduFigure 7: Comparison of kernels that compute Ax=bwhere Ais a matrix with dimension 25602560 , and xis a matrix with dimension 2560Batch size, where Batch size 2[1, 10] . All matrices are in half-precision format. 7.2 Deployment Optimized Matrix Multiply Kernels We have found that deploying our models using half-precision (16-bit) floating-point arithmetic does not measurably change recognition accuracy. Because deployment does not require any updates to

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

70

not measurably change recognition accuracy. Because deployment does not require any updates to the network weights, it is far less sensitive to numerical precision than training. Using half-precision arithmetic saves memory space and bandwidth, which is especially useful for deployment, since RNN evaluation is dominated by the cost of caching and streaming the weight matrices. As seen in Section 7.1, the batch size during deployment is much smaller than in training. We found that standard BLAS libraries are inefficient at this batch size. To overcome this, we wrote our own half-precision matrix-matrix multiply kernel. For 10 simultaneous streams over 90 percent of batches are for N4, a regime where the matrix multiply will be bandwidth bound. We store the A matrix transposed to maximize bandwidth by using the widest possible vector loads while avoiding transposition after loading. Each warp computes four rows of output for all Noutput columns. Note that forN4theBmatrix fits entirely in the L1 cache. This scheme achieves 90 percent of peak bandwidth for N4but starts to lose efficiency for larger Nas theBmatrix stops fitting into the L1 cache. Nonetheless, it continues to provide improved performance over existing libraries up to N= 10 .

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

71

L1 cache. Nonetheless, it continues to provide improved performance over existing libraries up to N= 10 . Figure 7 shows that our deployment kernel sustains a higher computational throughput than those from Nervana Systems [44] on the K1200 GPU, across the entire range of batch sizes that we use in deployment. Both our kernels and the Nervana kernels are significantly faster than NVIDIA CUBLAS version 7.0, more details are found here [20]. 7.3 Beam Search Performing the beam search involves repeated lookups in the n-gram language model, most of which translate to uncached reads from memory. The direct implementation of beam search means that each time-step dispatches one lookup per character for each beam. In Mandarin, this results in over 1M lookups per 40ms stride of speech data, which is too slow for deployment. To deal with this problem, we use a heuristic to further prune the beam search. Rather than considering all characters as viable additions to the beam, we only consider the fewest number of characters whose cumulative probability is at least p. In practice, we have found that p= 0.99 works well. Additionally, we limit ourselves to no more than 40 characters. This speeds up the Mandarin language model lookup time

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

72

ourselves to no more than 40 characters. This speeds up the Mandarin language model lookup time by a factor of 150x, and has a negligible effect on the CER (0.1-0.3% relative). 7.4 Results We can deploy our system at low latency and high throughput without sacrificing much accuracy. On a held-out set of 2000 utterances, our research system achieves 5.81 character error rate whereas the deployed system achieves 6.10 character error rate. This is only a 5% relative degradation for 22 the deployed system. In order to accomplish this, we employ a neural network architecture with low deployment latency, reduce the precision of our network to 16-bit, built a batching scheduler to more efficiently evaluate RNNs, and find a simple heuristic to reduce beam search cost. The model has five forward-only recurrent layers with 2560 hidden units, one row convolution layer (Section 3.7) with= 19 , and one fully-connected layer with 2560 hidden units. These techniques allow us to deploy Deep Speech at low cost to interactive applications. 8 Conclusion End-to-end deep learning presents the exciting opportunity to improve speech recognition systems continually with increases in data and computation. Indeed, our results show that, compared to the

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

73

8 Conclusion End-to-end deep learning presents the exciting opportunity to improve speech recognition systems continually with increases in data and computation. Indeed, our results show that, compared to the previous incarnation, Deep Speech has significantly closed the gap in transcription performance with human workers by leveraging more data and larger models. Further, since the approach is highly generic, we’ve shown that it can quickly be applied to new languages. Creating high-performing recognizers for two very different languages, English and Mandarin, required essentially no expert knowledge of the languages. Finally, we have also shown that this approach can be efficiently deployed by batching user requests together on a GPU server, paving the way to deliver end-to-end Deep Learning technologies to users. To achieve these results, we have explored various network architectures, finding several effective techniques: enhancements to numerical optimization through SortaGrad and Batch Normalization, evaluation of RNNs with larger strides with bigram outputs for English, searching through both bidirectional and unidirectional models. This exploration was powered by a well optimized, High Performance Computing inspired training system that allows us to train new, full-scale models on our large datasets in just a few days.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

74

Performance Computing inspired training system that allows us to train new, full-scale models on our large datasets in just a few days. Overall, we believe our results confirm and exemplify the value of end-to-end Deep Learning methods for speech recognition in several settings. In those cases where our system is not already comparable to humans, the difference has fallen rapidly, largely because of application-agnostic Deep Learning techniques. We believe these techniques will continue to scale, and thus conclude that the vision of a single speech system that outperforms humans in most scenarios is imminently achievable. Acknowledgments We are grateful to Baidu’s speech technology group for help with data preparation and useful conversations. We would like to thank Scott Gray, Amir Khosrowshahi and all of Nervana Systems for their excellent matrix multiply routines and useful discussions. We would also like to thank Natalia Gimelshein of NVIDIA for useful discussions and thoughts on implementing our fast deployment matrix multiply. References [1] O. Abdel-Hamid, A.-r. Mohamed, H. Jang, and G. Penn. Applying convolutional neural networks concepts to hybrid nn-hmm model for speech recognition. In ICASSP , 2012.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

75

[2] D. Bahdanau, K. Cho, and Y . Bengio. Neural machine translation by jointly learning to align and translate. InICLR , 2015. [3] D. Bahdanau, J. Chorowski, D. Serdyuk, P. Brakel, and Y . Bengio. End-to-end attention-based large vocabulary speech recognition. abs/1508.04395, 2015. http://arxiv.org/abs/1508.04395. [4] J. Barker, E. Marxer, Ricard Vincent, and S. Watanabe. The third ’CHiME’ speech separation and recognition challenge: Dataset, task and baselines. 2015. Submitted to IEEE 2015 Automatic Speech Recognition and Understanding Workshop (ASRU). [5] S. Baxter. Modern GPU. https://nvlabs.github.io/moderngpu/ . [6] Y . Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. In International Conference on Machine Learning , 2009. [7] H. Bourlard and N. Morgan. Connectionist Speech Recognition: A Hybrid Approach . Kluwer Academic Publishers, Norwell, MA, 1993. 23

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

76

[7] H. Bourlard and N. Morgan. Connectionist Speech Recognition: A Hybrid Approach . Kluwer Academic Publishers, Norwell, MA, 1993. 23 [8] W. Chan, N. Jaitly, Q. Le, and O. Vinyals. Listen, attend, and spell. abs/1508.01211, 2015. http://arxiv.org/abs/1508.01211. [9] S. Chetlur, C. Woolley, P. Vandermersch, J. Cohen, J. Tran, B. Catanzaro, and E. Shelhamer. cuDNN: Efficient primitives for deep learning. [10] T. Chilimbi, Y . Suzue, J. Apacible, and K. Kalyanaraman. Project adam: Building an efficient and scalable deep learning training system. In USENIX Symposium on Operating Systems Design and Implementation , 2014. [11] K. Cho, B. Van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP , 2014.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

77

[12] J. Chorowski, D. Bahdanau, K. Cho, and Y . Bengio. End-to-end continuous speech recognition using attention-based recurrent nn: First results. abs/1412.1602, 2015. http://arxiv.org/abs/1412.1602. [13] C. Cieri, D. Miller, and K. Walker. The Fisher corpus: a resource for the next generations of speech-totext. In LREC , volume 4, pages 69–71, 2004. [14] A. Coates, B. Carpenter, C. Case, S. Satheesh, B. Suresh, T. Wang, D. J. Wu, and A. Y . Ng. Text detection and character recognition in scene images with unsupervised feature learning. In International Conference on Document Analysis and Recognition , 2011. [15] A. Coates, B. Huval, T. Wang, D. J. Wu, A. Y . Ng, and B. Catanzaro. Deep learning with COTS HPC. In International Conference on Machine Learning , 2013. [16] G. Dahl, D. Yu, and L. Deng. Large vocabulary continuous speech recognition with context-dependent DBN-HMMs. In Proc. ICASSP , 2011.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

78

DBN-HMMs. In Proc. ICASSP , 2011. [17] G. Dahl, D. Yu, L. Deng, and A. Acero. Context-dependent pre-trained deep neural networks for large vocabulary speech recognition. IEEE Transactions on Audio, Speech, and Language Processing , 2011. [18] J. Dean, G. S. Corrado, R. Monga, K. Chen, M. Devin, Q. Le, M. Mao, M. Ranzato, A. Senior, P. Tucker, K. Yang, and A. Ng. Large scale distributed deep networks. In Advances in Neural Information Processing Systems 25 , 2012. [19] D. Ellis and N. Morgan. Size matters: An empirical study of neural network training for large vocabulary continuous speech recognition. In ICASSP , volume 2, pages 1013–1016. IEEE, 1999. [20] E. Elsen. Optimizing RNN performance. http://svail.github.io/rnn_perf . Accessed: 2015-11-24. [21] M. J. F. Gales, A. Ragni, H. Aldamarki, and C. Gautier. Support vector machines for noise robust ASR. InASRU , pages 205–2010, 2009.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

79

InASRU , pages 205–2010, 2009. [22] A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber. Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks. In ICML , pages 369–376. ACM, 2006. [23] A. Graves and N. Jaitly. Towards end-to-end speech recognition with recurrent neural networks. In ICML , 2014. [24] A. Graves, A.-r. Mohamed, and G. Hinton. Speech recognition with deep recurrent neural networks. In ICASSP , 2013. [25] H. H. Sak, A. Senior, and F. Beaufays. Long short-term memory recurrent neural network architectures for large scale acoustic modeling. In Interspeech , 2014. [26] A. Hannun, C. Case, J. Casper, B. Catanzaro, G. Diamos, E. Elsen, R. Prenger, S. Satheesh, S. Sengupta, A. Coates, and A. Y . Ng. Deep speech: Scaling up end-to-end speech recognition. 1412.5567, 2014. http://arxiv.org/abs/1412.5567.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

80

http://arxiv.org/abs/1412.5567. [27] A. Y . Hannun, A. L. Maas, D. Jurafsky, and A. Y . Ng. First-pass large vocabulary continuous speech recognition using bi-directional recurrent DNNs. abs/1408.2873, 2014. http://arxiv.org/abs/1408.2873. [28] K. Heafield, I. Pouzyrevsky, J. H. Clark, and P. Koehn. Scalable modified Kneser-Ney language model estimation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics , Sofia, Bulgaria, 8 2013. [29] G. Hinton, L. Deng, D. Yu, G. Dahl, A. Mohamed, N. Jaitly, A. Senior, V . Vanhoucke, P. Nguyen, T. Sainath, and B. Kingsbury. Deep neural networks for acoustic modeling in speech recognition. IEEE Signal Processing Magazine , 29(November):82–97, 2012.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

81

T. Sainath, and B. Kingsbury. Deep neural networks for acoustic modeling in speech recognition. IEEE Signal Processing Magazine , 29(November):82–97, 2012. [30] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation , 9(8):1735—1780, 1997. [31] N. Jaitly and G. Hinton. V ocal tract length perturbation (VTLP) improves speech recognition. In ICML Workshop on Deep Learning for Audio, Speech, and Language Processing , 2013. [32] R. Jozefowicz, W. Zaremba, and I. Sutskever. An empirical exploration of recurrent network architectures. InICML , 2015. 24 [33] O. Kapralova, J. Alex, E. Weinstein, P. Moreno, and O. Siohan. A big data approach to acoustic model training corpus selection. In Interspeech , 2014. [34] K. C. Knowlton. A fast storage allocator. Commun. ACM , 8(10):623–624, Oct. 1965.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

82

[34] K. C. Knowlton. A fast storage allocator. Commun. ACM , 8(10):623–624, Oct. 1965. [35] T. Ko, V . Peddinti, D. Povey, and S. Khudanpur. Audio augmentation for speech recognition. In Interspeech , 2015. [36] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems 25 , pages 1106–1114, 2012. [37] C. Laurent, G. Pereyra, P. Brakel, Y . Zhang, and Y . Bengio. Batch normalized recurrent neural networks. abs/1510.01378, 2015. http://arxiv.org/abs/1510.01378. [38] Q. Le, M. Ranzato, R. Monga, M. Devin, K. Chen, G. Corrado, J. Dean, and A. Ng. Building high-level features using large scale unsupervised learning. In International Conference on Machine Learning , 2012. [39] Y . LeCun, F. J. Huang, and L. Bottou. Learning methods for generic object recognition with invariance

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

83

[39] Y . LeCun, F. J. Huang, and L. Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Computer Vision and Pattern Recognition , volume 2, pages 97–104, 2004. [40] A. Maas, Z. Xie, D. Jurafsky, and A. Ng. Lexicon-free conversational speech recognition with neural networks. In NAACL , 2015. [41] Y . Miao, M. Gowayyed, and F. Metz. EESEN: End-to-end speech recognition using deep rnn models and wfst-based decoding. In ASRU , 2015. [42] A. Mohamed, G. Dahl, and G. Hinton. Acoustic modeling using deep belief networks. IEEE Transactions on Audio, Speech, and Language Processing , (99), 2011. [43] A. S. N. Jaitly, P. Nguyen and V . Vanhoucke. Application of pretrained deep neural networks to large vocabulary speech recognition. In Interspeech , 2012. [44] Nervana Systems. Nervana GPU. https://github.com/NervanaSystems/nervanagpu . Accessed: 2015-11-06.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

84

[44] Nervana Systems. Nervana GPU. https://github.com/NervanaSystems/nervanagpu . Accessed: 2015-11-06. [45] J. Niu, L. Xie, L. Jia, and N. Hu. Context-dependent deep neural networks for commercial mandarin speech recognition applications. In APSIPA , 2013. [46] V . Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an asr corpus based on public domain audio books. In ICASSP , 2015. [47] R. Pascanu, T. Mikolov, and Y . Bengio. On the difficulty of training recurrent neural networks. abs/1211.5063, 2012. http://arxiv.org/abs/1211.5063. [48] P. Patarasuk and X. Yuan. Bandwidth optimal all-reduce algorithms for clusters of workstations. J. Parallel Distrib. Comput. , 69(2):117–124, Feb. 2009. [49] R. Raina, A. Madhavan, and A. Ng. Large-scale deep unsupervised learning using graphics processors. In26th International Conference on Machine Learning , 2009.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

85

In26th International Conference on Machine Learning , 2009. [50] S. Renals, N. Morgan, H. Bourlard, M. Cohen, and H. Franco. Connectionist probability estimators in HMM speech recognition. IEEE Transactions on Speech and Audio Processing , 2(1):161–174, 1994. [51] T. Robinson, M. Hochberg, and S. Renals. The use of recurrent neural networks in continuous speech recognition. pages 253–258, 1996. [52] T. Sainath, O. Vinyals, A. Senior, and H. Sak. Convolutional, long short-term memory, fully connected deep neural networks. In ICASSP , 2015. [53] T. N. Sainath, A. rahman Mohamed, B. Kingsbury, and B. Ramabhadran. Deep convolutional neural networks for LVCSR. In ICASSP , 2013. [54] H. Sak, A. Senior, K. Rao, and F. Beaufays. Fast and accurate recurrent neural network acoustic models for speech recognition. abs/1507.06947, 2015. http://arxiv.org/abs/1507.06947.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

86

for speech recognition. abs/1507.06947, 2015. http://arxiv.org/abs/1507.06947. [55] H. Sak, O. Vinyals, G. Heigold, A. Senior, E. McDermott, R. Monga, and M. Mao. Sequence discriminative distributed training of long shortterm memory recurrent neural networks. In Interspeech , 2014. [56] B. Sapp, A. Saxena, and A. Ng. A fast data collection and augmentation procedure for object recognition. InAAAI Twenty-Third Conference on Artificial Intelligence , 2008. [57] M. Schuster and K. K. Paliwal. Bidirectional recurrent neural networks. IEEE Transactions on Signal Processing , 45(11):2673–2681, 1997. [58] F. Seide, G. Li, and D. Yu. Conversational speech transcription using context-dependent deep neural networks. In Interspeech , pages 437–440, 2011. [59] J. Shan, G. Wu, Z. Hu, X. Tang, M. Jansche, and P. Moreno. Search by voice in mandarin chinese. In Interspeech , 2010.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

87

Interspeech , 2010. [60] H. Soltau, G. Saon, and T. Sainath. Joint training of convolutional and non-convolutional neural networks. InICASSP , 2014. 25 [61] I. Sutskever, J. Martens, G. Dahl, and G. Hinton. On the importance of momentum and initialization in deep learning. In 30th International Conference on Machine Learning , 2013. [62] I. Sutskever, O. Vinyals, and Q. V . Le. Sequence to sequence learning with neural networks. 2014. http://arxiv.org/abs/1409.3215. [63] C. Szegedy and S. Ioffe. Batch normalization: Accelerating deep network training by reducing internal covariate shift. abs/1502.03167, 2015. http://arxiv.org/abs/1502.03167. [64] C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich. Going deeper with convolutions. 2014.

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

88

[65] R. Thakur and R. Rabenseifner. Optimization of collective communication operations in mpich. International Journal of High Performance Computing Applications , 19:49–66, 2005. [66] K. Vesely, A. Ghoshal, L. Burget, and D. Povey. Sequence-discriminative training of deep neural networks. In Interspeech , 2013. [67] A. Waibel, T. Hanazawa, G. Hinton, K. Shikano, and K. Lang. Phoneme recognition using time-delay neural networks,â ˘A˙I acoustics speech and signal processing. IEEE Transactions on Acoustics, Speech and Signal Processing , 37(3):328–339, 1989. [68] R. Williams and J. Peng. An efficient gradient-based algorithm for online training of recurrent network trajectories. Neural computation , 2:490–501, 1990. [69] T. Yoshioka, N. Ito, M. Delcroix, A. Ogawa, K. Kinoshita, M. F. C. Yu, W. J. Fabian, M. Espi, T. Higuchi,

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

89

S. Araki, and T. Nakatani. The ntt chime-3 system: Advances in speech enhancement and recognition for mobile multi-microphone devices. In IEEE ASRU , 2015. [70] W. Zaremba and I. Sutskever. Learning to execute. abs/1410.4615, 2014. http://arxiv.org/abs/1410.4615. A Scalability improvements In this section, we discuss some of our scalability improvements in more detail. A.1 Node and cluster architecture The software stack runs on a compute dense node built from 2Intel CPUs and 8NVIDIA Titan X GPUs, with peak single-precision computational throughput of 53teraFLOP/second. Each node also has 384GB of CPU memory and an 8TB storage volume built from two 4 TB hard disks in RAID-0 configuration. We use the CPU memory to cache our input data so that we are not directly exposed to the low bandwidth and high latency of spinning disks. We replicate our English and Mandarin datasets on each node’s local hard disk. This allows us to use our network only for weight updates and avoids having to rely on centralized file servers. GPUGPUGPUGPUGPUGPUGPUGPUPLXPLXPLXCPUCPUPLX

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

90

updates and avoids having to rely on centralized file servers. GPUGPUGPUGPUGPUGPUGPUGPUPLXPLXPLXCPUCPUPLX Figure 8: Schematic of our training node where PLX indicates a PCI switch and the dotted box includes all devices that are connected by the same PCI root complex. Figure 8 shows a schematic diagram of one our nodes, where all devices connected by the same PCI root complex are encapsulated in a dotted box. We have tried to maximize the number of GPUs within the root complex for faster communication between GPUs using GPUDirect. This allows us to use an efficient communication mechanism to transfer gradient matrices between GPUs. 26 All the nodes in our cluster are connected through Fourteen Data Rate (FDR) Infiniband which is primarily used for gradient transfer during back-propagation. A.2 GPU Implementation of CTC Loss Function The CTC loss function that we use to train our models has two passes: forward and backward, and the gradient computation involves element-wise addition of two matrices, and, generated during the forward and backward passes respectively. Finally, we sum the gradients using the character in

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

91

the gradient computation involves element-wise addition of two matrices, and, generated during the forward and backward passes respectively. Finally, we sum the gradients using the character in the utterance label as the key, to generate one gradient per character. These gradients are then backpropagated through the network. The input to the CTC loss function are probabilities calculated by the softmax function which can be very small, so we compute in logprobability space for better numerical stability. The forward pass of the CTC algorithm calculates the matrix, which has Srows and Tcolumns, where S= 2(L+ 1) . The variable Lis the number of characters in the label and Tis the number of time-steps in the utterance. Our CPU-based implementation of the CTC algorithm assigns one thread to each utterance label in a minibatch, performing the CTC calculation for the utterances in parallel. Each thread calculates the relevant entries of the matrix sequentially. This is inefficient for two reasons. Firstly, since the remainder of our network is computed on the GPU, the output of the softmax function has to be copied to the CPU for CTC calculation. The gradient matrices from the CTC function then has to be copied back to the GPU for backpropagation. For languages like Mandarin

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

92

function then has to be copied back to the GPU for backpropagation. For languages like Mandarin with large character sets, these matrices have hundreds of millions of entries, making this copy expensive. Furthermore, we need as much interconnect bandwidth as possible for synchronizing the gradient updates with data parallelism, so this copy incurs a substantial opportunity cost. Secondly, although entries in each column of the matrix can be computed in parallel, the number of entries to calculate in each column depends both on the column and the number of repeated characters in the utterance label. Due to this complexity, the CPU implementation does not use SIMD parallelism optimally, making the computation inefficient. We wrote a GPU-based implementation of CTC in order to overcome these two problems. The key insight behind our implementation is that we can compute all elements in each column of the  matrix, rather than just the valid entries. If we do so, Figure 9 shows that invalid elements either contain a finite garbage value ( G), or1 (I), when we use a special summation function that adds probabilities in logspace that discards inputs that are 1. This summation is shown in Figure 9 where arrows incident on a circle are inputs and the result is stored in the circle. However, when we

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

93

where arrows incident on a circle are inputs and the result is stored in the circle. However, when we compute the final gradient by element-wise summing and, all finite garbage values will be added with a corresponding 1 value from the other matrix, which results in 1, effectively ignoring the garbage value and computing the correct result. One important observation is that this element-wise sum ofandis a simple sum and does not use our summation function. To compute the gradient, we take each column of the matrix generated from element-wise addition ofandmatrices, and do a key-value reduction using the character as key, using the ModernGPU library [5]. This means elements of the column corresponding to the same character will sum up their values. In the example shown in Figure 9, the blank character, B, is the only repeated character and at some columns, say for t= 1 oft= 2, both valid elements (gray) and 1 correspond to it. Since our summation function in logspace effectively ignores the 1 elements, only the valid elements are combined in the reduction. In our GPU implementation, we map each utterance in the minibatch to a CUDA thread block .

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

94

elements are combined in the reduction. In our GPU implementation, we map each utterance in the minibatch to a CUDA thread block . Since there are no dependencies between the elements of a column, all of them can be computed in parallel by the threads in a threadblock. There are dependencies between columns, since the column corresponding to time-step t+ 1cannot be computed before the column corresponding to time-step t. The reverse happens when computing the matrix, when column corresponding to time-step tcannot be computed before the column corresponding to time-step t+ 1. Thus, in both cases, columns are processed sequentially by the thread block. Mapping the forward and backward passes to corresponding CUDA kernels is straightforward since there are no data dependencies between elements of a column. The kernel that does the backward pass also computes the gradient. However, since the gradients must be summed up based on the label 27 IIIIATCIIIIIGGGGGGGG G1234T-3T-2T-1T GGGGATCGGGGGIIIIIIII I↵BBBBBBBBFigure 9: Forward and backward pass for GPU implementation of CTC. Gray circles contain valid values,

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

95

GGGGATCGGGGGIIIIIIII I↵BBBBBBBBFigure 9: Forward and backward pass for GPU implementation of CTC. Gray circles contain valid values, circle with Icontain1 and circle with Gcontain garbage values that are finite. Bstand for the blank character that the CTC algorithm adds to the input utterance label. Column labels on top show different time-steps going from1toT. values, with each character as key, we must deal with data dependencies due to repeated characters in an utterance label. For languages with small character sets like English, this happens with high probability. Even if there are no repeated characters, the CTC algorithm adds L+ 1blank characters to the utterance label. We solve this problem by performing a key-value sort, where the keys are the characters in the utterance label, and the values are the indices of each character in the utterance. After sorting, all occurrences of a given character are arranged in contiguous segments. We only need to do the sort once for each utterance. The indices generated by the sort are then used to sequentially sum up the gradients for each character. This sum is done once per column and in parallel over all characters in the utterance. Amortizing the cost of key-value sort over Tcolumns is

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1512.02595

96

parallel over all characters in the utterance. Amortizing the cost of key-value sort over Tcolumns is a key insight that makes the gradient calculation fast. Our GPU implementation uses fast shared memory and registers to achieve high performance when performing this task. Both forward and backward kernels store the matrix in shared memory . Since shared memory is a limited resource, it is not possible to store the entire matrix. However, as we go backward in time, we only need to keep one column of the matrix as we compute the gradient, adding element-wise the column of the matrix with the corresponding column of the matrix. Due to on-chip memory space constraints, we read the output of the softmax function directly from off-chip global memory . Due to inaccuracies in floating-point arithmetic, especially in transcendental functions, our GPU and CPU implementation are not bit-wise identical. This is not an impediment in practice, since both implementations train models equally well when coupled with the technique of sorting utterances by length mentioned in Section 3.3. 28

1512.02595

Deep Speech 2: End-to-End Speech Recognition in English and Mandarin

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech--two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, resulting in a 7x speedup over our previous system. Because of this efficiency, experiments that previously took weeks now run in days. This enables us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

http://arxiv.org/pdf/1512.02595

cs.CL

20151208

20151208

1603.09025

0

Published as a conference paper at ICLR 2017 RECURRENT BATCH NORMALIZATION Tim Cooijmans, Nicolas Ballas, César Laurent, Ça ˘glar Gülçehre & Aaron Courville MILA - Université de Montréal firstname.lastname@umontreal.ca ABSTRACT We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization. 1 I NTRODUCTION Recurrent neural network architectures such as LSTM (Hochreiter & Schmidhuber, 1997) and GRU (Cho et al., 2014) have recently exhibited state-of-the-art performance on a wide range of

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

1

GRU (Cho et al., 2014) have recently exhibited state-of-the-art performance on a wide range of complex sequential problems including speech recognition Amodei et al. (2015), machine translation (Bahdanau et al., 2015) and image and video captioning (Xu et al., 2015; Yao et al., 2015). Top-performing models, however, are based on very high-capacity networks that are computationally intensive and costly to train. Effective optimization of recurrent neural networks is thus an active area of study (Pascanu et al., 2012; Martens & Sutskever, 2011; Ollivier, 2013). It is well-known that for deep feed-forward neural networks, covariate shift (Shimodaira, 2000; Ioffe & Szegedy, 2015) degrades the efficiency of training. Covariate shift is a change in the distribution of the inputs to a model. This occurs continuously during training of feed-forward neural networks, where changing the parameters of a layer affects the distribution of the inputs to all layers above it. As a result, the upper layers are continually adapting to the shifting input distribution and unable to learn effectively. This internal covariate shift (Ioffe & Szegedy, 2015) may play an especially important role in recurrent neural networks, which resemble very deep feed-forward networks.

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

2

important role in recurrent neural networks, which resemble very deep feed-forward networks. Batch normalization (Ioffe & Szegedy, 2015) is a recently proposed technique for controlling the distributions of feed-forward neural network activations, thereby reducing internal covariate shift. It involves standardizing the activations going into each layer, enforcing their means and variances to be invariant to changes in the parameters of the underlying layers. This effectively decouples each layer’s parameters from those of other layers, leading to a better-conditioned optimization problem. Indeed, deep neural networks trained with batch normalization converge significantly faster and generalize better. Although batch normalization has demonstrated significant training speed-ups and generalization benefits in feed-forward networks, it is proven to be difficult to apply in recurrent architectures (Laurent et al., 2016; Amodei et al., 2015). It has found limited use in stacked RNNs, where the normalization is applied “vertically”, i.e. to the input of each RNN, but not “horizontally” between timesteps. RNNs are deeper in the time direction, and as such batch normalization would be most

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

3

timesteps. RNNs are deeper in the time direction, and as such batch normalization would be most beneficial when applied horizontally. However, Laurent et al. (2016) hypothesized that applying batch normalization in this way hurts training because of exploding gradients due to repeated rescaling. Our findings run counter to this hypothesis. We show that it is both possible and highly beneficial to apply batch normalization in the hidden-to-hidden transition of recurrent models. In particular, we describe a reparameterization of LSTM (Section 3) that involves batch normalization and demonstrate that it is easier to optimize and generalizes better. In addition, we empirically analyze the 1arXiv:1603.09025v5 [cs.LG] 28 Feb 2017 Published as a conference paper at ICLR 2017 gradient backpropagation and show that proper initialization of the batch normalization parameters is crucial to avoiding vanishing gradient (Section 4). We evaluate our proposal on several sequential problems and show (Section 5) that our LSTM reparameterization consistently outperforms the LSTM baseline across tasks, in terms of both time to convergence and performance. Liao & Poggio (2016) simultaneously investigated batch normalization in recurrent neural networks,

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

4

LSTM baseline across tasks, in terms of both time to convergence and performance. Liao & Poggio (2016) simultaneously investigated batch normalization in recurrent neural networks, albeit only for very short sequences (10 steps). Ba et al. (2016) independently developed a variant of batch normalization that is also applicable to recurrent neural networks and delivers similar improvements as our method. 2 P REREQUISITES 2.1 LSTM Long Short-Term Memory (LSTM) networks are an instance of a more general class of recurrent neural networks (RNNs), which we review briefly in this paper. Given an input sequence X= (x1;x2;:::;xT), an RNN defines a sequence of hidden states htaccording to ht=(Whht1+Wxxt+b); (1) where Wh2Rdhdh;Wx2Rdxdh;b2Rdhand the initial state h02Rdhare model parameters. A popular choice for the activation function (
)istanh . RNNs are popular in sequence modeling thanks to their natural ability to process variable-length

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

5

RNNs are popular in sequence modeling thanks to their natural ability to process variable-length sequences. However, training RNNs using first-order stochastic gradient descent (SGD) is notoriously difficult due to the well-known problem of exploding/vanishing gradients (Bengio et al., 1994; Hochreiter, 1991; Pascanu et al., 2012). Gradient vanishing occurs when states htare not influenced by small changes in much earlier states h,t, preventing learning of long-term dependencies in the input data. Although learning long-term dependencies is fundamentally difficult (Bengio et al., 1994), its effects can be mitigated through architectural variations such as LSTM (Hochreiter & Schmidhuber, 1997), GRU (Cho et al., 2014) and iRNN/uRNN (Le et al., 2015; Arjovsky et al., 2015). In what follows, we focus on the LSTM architecture (Hochreiter & Schmidhuber, 1997) with recurrent transition given by 0 BB@~ft ~it ~ot ~gt1 CCA=Whht1+Wxxt+b (2)

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

6

0 BB@~ft ~it ~ot ~gt1 CCA=Whht1+Wxxt+b (2) ct=(~ft)ct1+(~it)tanh( ~gt) (3) ht=(~ot)tanh(ct); (4) where Wh2Rdh4dh;WxRdx4dh;b2R4dhand the initial states h02Rdh;c02Rdhare model parameters. is the logistic sigmoid function, and the operator denotes the Hadamard product. The LSTM differs from simple RNNs in that it has an additional memory cellctwhose update is nearly linear which allows the gradient to flow back through time more easily. In addition, unlike the RNN which overwrites its content at each timestep, the update of the LSTM cell is regulated by a set of gates. The forget gate ftdetermines the extent to which information is carried over from the previous timestep, and the input gate itcontrols the flow of information from the current input xt. The output gate otallows the model to read from the cell. This carefully controlled interaction with

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

7

The output gate otallows the model to read from the cell. This carefully controlled interaction with the cell is what allows the LSTM to robustly retain information for long periods of time. 2.2 B ATCH NORMALIZATION Covariate shift (Shimodaira, 2000) is a phenomenon in machine learning where the features presented to a model change in distribution. In order for learning to succeed in the presence of covariate shift, the model’s parameters must be adjusted not just to learn the concept at hand but also to adapt to the changing distribution of the inputs. In deep neural networks, this problem manifests as 2 Published as a conference paper at ICLR 2017 internal covariate shift (Ioffe & Szegedy, 2015), where changing the parameters of a layer affects the distribution of the inputs to all layers above it. Batch Normalization (Ioffe & Szegedy, 2015) is a recently proposed network reparameterization which aims to reduce internal covariate shift. It does so by standardizing the activations using empirical estimates of their means and standard deviations. However, it does not decorrelate the activations due to the computationally costly matrix inversion. The batch normalizing transform is as follows: BN(h; ;) =+ hbE[h]q

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

8

as follows: BN(h; ;) =+ hbE[h]q dVar[h] +(5) where h2Rdis the vector of (pre)activations to be normalized, 2Rd;2Rdare model parameters that determine the mean and standard deviation of the normalized activation, and 2R is a regularization hyperparameter. The division should be understood to proceed elementwise. At training time, the statistics E[h]andVar[h]are estimated by the sample mean and sample variance of the current minibatch. This allows for backpropagation through the statistics, preserving the convergence properties of stochastic gradient descent. During inference, the statistics are typically estimated based on the entire training set, so as to produce a deterministic prediction. 3 B ATCH -NORMALIZED LSTM This section introduces a reparameterization of LSTM that takes advantage of batch normalization. Contrary to Laurent et al. (2016); Amodei et al. (2015), we leverage batch normalization in both the input-to-hidden andthe hidden-to-hidden transformations. We introduce the batch-normalizing transform BN(
; ;)into the LSTM as follows: 0 BB@~ft

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

9

transform BN(
; ;)into the LSTM as follows: 0 BB@~ft ~it ~ot ~gt1 CCA= BN( Whht1; h;h) + BN( Wxxt; x;x) +b (6) ct=(~ft)ct1+(~it)tanh( ~gt) (7) ht=(~ot)tanh(BN( ct; c;c)) (8) In our formulation, we normalize the recurrent term Whht1and the input term Wxxtseparately. Normalizing these terms individually gives the model better control over the relative contribution of the terms using the hand xparameters. We set h=x=0to avoid unnecessary redundancy, instead relying on the pre-existing parameter vector bto account for both biases. In order to leave the LSTM dynamics intact and preserve the gradient flow through ct, we do not apply batch normalization in the cell update. The batch normalization transform relies on batch statistics to standardize the LSTM activations. It would seem natural to share the statistics that are used for normalization across time, just as recurrent

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

10

The batch normalization transform relies on batch statistics to standardize the LSTM activations. It would seem natural to share the statistics that are used for normalization across time, just as recurrent neural networks share their parameters over time. However, we find that simply averaging statistics over time severely degrades performance. Although LSTM activations do converge to a stationary distribution, we observe that their statistics during the initial transient differ significantly (see Figure 5 in Appendix A). Consequently, we recommend using separate statistics for each timestep to preserve information of the initial transient phase in the activations.1 Generalizing the model to sequences longer than those seen during training is straightforward thanks to the rapid convergence of the activations to their steady-state distributions (cf. Figure 5). For our experiments we estimate the population statistics separately for each timestep 1;:::;Tmax where 1Note that we separate only the statistics over time and not the andparameters. 3 Published as a conference paper at ICLR 2017 Tmax is the length of the longest training sequence. When at test time we need to generalize beyond Tmax, we use the population statistic of time Tmax for all time steps beyond it. During training we estimate the statistics across the minibatch, independently for each timestep. At

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

11

Tmax, we use the population statistic of time Tmax for all time steps beyond it. During training we estimate the statistics across the minibatch, independently for each timestep. At test time we use estimates obtained by averaging the minibatch estimates over the training set. 4 I NITIALIZING FOR GRADIENT FLOW Although batch normalization allows for easy control of the pre-activation variance through the parameters, common practice is to normalize to unit variance. We suspect that the previous difficulties with recurrent batch normalization reported in Laurent et al. (2016); Amodei et al. (2015) are largely due to improper initialization of the batch normalization parameters, and in particular. In this section we demonstrate the impact of on gradient flow. 0 100 200 300 400 500 600 700 800 t10-2610-2410-2210-2010-1810-1610-1410-1210-1010-810-610-410-2100||dloss/dh_t||_2RNN gradient propagation gamma=0.10 gamma=0.20 gamma=0.30 gamma=0.40 gamma=0.50 gamma=0.60 gamma=0.70 gamma=0.80

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

12

gamma=0.40 gamma=0.50 gamma=0.60 gamma=0.70 gamma=0.80 gamma=0.90 gamma=1.00 (a) We visualize the gradient flow through a batchnormalized tanh RNN as a function of . High variance causes vanishing gradient. 0.0 0.2 0.4 0.6 0.8 1.0 input standard deviation0.00.20.40.60.81.0expected derivative (and IQR range)derivative through tanh(b) We show the empirical expected derivative and interquartile range of tanh nonlinearity as a function of input variance. High variance causes saturation, which decreases the expected derivative. Figure 1: Influence of pre-activation variance on gradient propagation. In Figure 1(a), we show how the pre-activation variance impacts gradient propagation in a simple RNN on the sequential MNIST task described in Section 5.1. Since backpropagation operates in reverse, the plot is best read from right to left. The quantity plotted is the norm of the gradient of the loss with respect to the hidden state at different time steps. For large values of , the norm

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

13

of the loss with respect to the hidden state at different time steps. For large values of , the norm quickly goes to zero as gradient is propagated back in time. For small values of the norm is nearly constant. To demonstrate what we think is the cause of this vanishing, we drew samples xfrom a set of centered Gaussian distributions with standard deviation ranging from 0 to 1, and computed the derivative tanh0(x) = 1tanh2(x)2[0;1]for each. Figure 1(b) shows the empirical distribution of the derivative as a function of standard deviation. When the input standard deviation is low, the input tends to be close to the origin where the derivative is close to 1. As the standard deviation increases, the expected derivative decreases as the input is more likely to be in the saturation regime. At unit standard deviation, the expected derivative is much smaller than 1. We conjecture that this is what causes the gradient to vanish, and recommend initializing to a small value. In our trials we found that values of 0.01 or lower caused instabilities during training. Our choice of 0.1 seems to work well across different tasks. 5 E XPERIMENTS

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

14

Our choice of 0.1 seems to work well across different tasks. 5 E XPERIMENTS This section presents an empirical evaluation of the proposed batch-normalized LSTM on four different tasks. Note that for all the experiments, we initialize the batch normalization scale and shift parameters andto0:1and0respectively. 4 Published as a conference paper at ICLR 2017 0 20000 40000 60000 80000 100000 Training Iteration0.00.20.40.60.81.0AccuracyPixel-by-Pixel MNIST (Validation Set) lstm bn_lstm 0 20000 40000 60000 80000 100000 Training Iteration0.10.20.30.40.50.60.70.80.91.0AccuracyPixel-by-Pixel Permuted-MNIST (Validation Set) lstm bn_lstm Figure 2: Accuracy on the validation set for the pixel by pixel MNIST classification tasks. The batch-normalized LSTM is able to converge faster relatively to a baseline LSTM. Batch-normalized LSTM also shows some improve generalization on the permuted sequential MNIST that require to preserve long-term memory information. 5.1 S EQUENTIAL MNIST

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

15

LSTM also shows some improve generalization on the permuted sequential MNIST that require to preserve long-term memory information. 5.1 S EQUENTIAL MNIST We evaluate our batch-normalized LSTM on a sequential version of the MNIST classification task (Le et al., 2015). The model processes each image one pixel at a time and finally predicts the label. We consider both sequential MNIST tasks, MNIST and permuted MNIST ( pMNIST). In MNIST, the pixels are processed in scanline order. In pMNIST the pixels are processed in a fixed random order. Our baseline consists of an LSTM with 100 hidden units, with a softmax classifier to produce a prediction from the final hidden state. We use orthogonal initialization for all weight matrices, except for the hidden-to-hidden weight matrix which we initialize to be the identity matrix, as this yields better generalization performance on this task for both models. The model is trained using RMSProp (Tieleman & Hinton, 2012) with learning rate of 103and0:9momentum. We apply gradient clipping at 1 to avoid exploding gradients.

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

16

gradient clipping at 1 to avoid exploding gradients. The in-order MNIST task poses a unique problem for our model: the input for the first hundred or so timesteps is constant across examples since the upper pixels are almost always black. This causes the variance of the hidden states to be exactly zero for a long period of time. Normalizing these zerovariance activations involves dividing zero by a small number at many timesteps, which does not affect the forward-propagated activations but causes the back-propagated gradient to explode. We work around this by adding Gaussian noise to the initial hidden states. Although the normalization amplifies the noise to signal level, we find that it does not hurt performance compared to datadependent ways of initializing the hidden states. Model MNIST pMNIST TANH-RNN (Le et al., 2015) 35.0 35.0 iRNN (Le et al., 2015) 97.0 82.0 uRNN (Arjovsky et al., 2015) 95.1 91.4 sTANH-RNN (Zhang et al., 2016) 98.1 94.0 LSTM (ours) 98.9 90.2 BN-LSTM (ours) 99.0 95.4

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

17

LSTM (ours) 98.9 90.2 BN-LSTM (ours) 99.0 95.4 Table 1: Accuracy obtained on the test set for the pixel by pixel MNIST classification tasks In Figure 2 we show the validation accuracy while training for both LSTM and batch-normalized LSTM (BN-LSTM). BN-LSTM converges faster than LSTM on both tasks. Additionally, we observe that BN-LSTM generalizes significantly better on pMNIST. It has been highlighted in Arjovsky et al. (2015) that pMNIST contains many longer term dependencies across pixels than in the original pixel ordering, where a lot of structure is local. A recurrent network therefore needs to 5 Published as a conference paper at ICLR 2017 Model Penn Treebank LSTM (Graves, 2013) 1.262 HF-MRNN (Mikolov et al., 2012) 1.41 Norm-stabilized LSTM (Krueger & Memisevic, 2016) 1.39 ME n-gram (Mikolov et al., 2012) 1.37 LSTM (ours) 1.38 BN-LSTM (ours) 1.32 Zoneout (Krueger et al., 2016) 1.27

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

18

LSTM (ours) 1.38 BN-LSTM (ours) 1.32 Zoneout (Krueger et al., 2016) 1.27 HM-LSTM (Chung et al., 2016) 1.24 HyperNetworks (Ha et al., 2016) 1.22 Table 2: Bits-per-character on the Penn Treebank test sequence. characterize dependencies across varying time scales in order to solve this task. Our results suggest that BN-LSTM is better able to capture these long-term dependencies. Table 1 reports the test set accuracy of the early stop model for LSTM and BN-LSTM using the population statistics. Recurrent batch normalization leads to a better test score, especially for pMNIST where models have to leverage long-term temporal depencies. In addition, Table 1 shows that our batch-normalized LSTM achieves state of the art on both MNIST and pMNIST. 5.2 C HARACTER -LEVEL PENN TREEBANK We evaluate our model on the task of character-level language modeling on the Penn Treebank corpus (Marcus et al., 1993) according to the train/valid/test partition of Mikolov et al. (2012). For training, we segment the training sequence into examples of length 100. The training sequence does

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

19

training, we segment the training sequence into examples of length 100. The training sequence does not cleanly divide by 100, so for each epoch we randomly crop a subsequence that does and segment that instead. Our baseline is an LSTM with 1000 units, trained to predict the next character using a softmax classifier on the hidden state ht. We use stochastic gradient descent on minibatches of size 64, with gradient clipping at 1.0 and step rule determined by Adam (Kingma & Ba, 2014) with learning rate 0.002. We use orthogonal initialization for all weight matrices. The setup for the batch-normalized LSTM is the same in all respects except for the introduction of batch normalization as detailed in 3. We show the learning curves in Figure 3(a). BN-LSTM converges faster and generalizes better than the LSTM baseline. Figure 3(b) shows the generalization of our model to longer sequences. We observe that using the population statistics improves generalization performance, which confirms that repeating the last population statistic (cf. Section 3) is a viable strategy. In table 2 we report the performance of our best models (early-stopped on validation performance) on the Penn Treebank

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

20

the performance of our best models (early-stopped on validation performance) on the Penn Treebank test sequence. Follow up works havd since improved the state of the art (Krueger et al., 2016; Chung et al., 2016; Ha et al., 2016). 5.3 T EXT8 We evaluate our model on a second character-level language modeling task on the much larger text8 dataset (Mahoney, 2009). This dataset is derived from Wikipedia and consists of a sequence of 100M characters including only alphabetical characters and spaces. We follow Mikolov et al. (2012); Zhang et al. (2016) and use the first 90M characters for training, the next 5M for validation and the final 5M characters for testing. We train on nonoverlapping sequences of length 180. Both our baseline and batch-normalized models are LSTMs with 2000 units, trained to predict the next character using a softmax classifier on the hidden state ht. We use stochastic gradient descent on minibatches of size 128, with gradient clipping at 1.0 and step rule determined by Adam (Kingma & Ba, 2014) with learning rate 0.001. All weight matrices were initialized to be orthogonal. 6 Published as a conference paper at ICLR 2017

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

21

& Ba, 2014) with learning rate 0.001. All weight matrices were initialized to be orthogonal. 6 Published as a conference paper at ICLR 2017 We early-stop on validation performance and report the test performance of the resulting model in table 3. We observe that BN-LSTM obtains a significant performance improvement over the LSTM baseline. Chung et al. (2016) has since improved on our performance. Model text8 td-LSTM (Zhang et al., 2016) 1.63 HF-MRNN (Mikolov et al., 2012) 1.54 skipping RNN (Pachitariu & Sahani, 2013) 1.48 LSTM (ours) 1.43 BN-LSTM (ours) 1.36 HM-LSTM (Chung et al., 2016) 1.29 Table 3: Bits-per-character on the text8 test sequence. 5.4 T EACHING MACHINES TO READ AND COMPREHEND Recently, Hermann et al. (2015) introduced a set of challenging benchmarks for natural language processing, along with neural network architectures to address them. The tasks involve reading

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

22

Recently, Hermann et al. (2015) introduced a set of challenging benchmarks for natural language processing, along with neural network architectures to address them. The tasks involve reading real news articles and answering questions about their content. Their principal model, the Attentive Reader, is a recurrent neural network that invokes an attention mechanism to locate relevant information in the document. Such models are notoriously hard to optimize and yet increasingly popular. To demonstrate the generality and practical applicability of our proposal, we apply batch normalization in the Attentive Reader model and show that this drastically improves training. We evaluate several variants. The first variant, referred to as BN-LSTM, consists of the vanilla Attentive Reader model with the LSTM simply replaced by our BN-LSTM reparameterization. The second variant, termed BN-everywhere, is exactly like the first, except that we also introduce batch normalization into the attention computations, normalizing each term going into the tanh nonlinearities. Our third variant, BN-e*, is like BN-everywhere, but improved to more carefully handle variablelength sequences. Throughout this experiment we followed the common practice of padding each batch of variable-length data with zeros. However, this biases the batch mean and variance of xt

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

23

batch of variable-length data with zeros. However, this biases the batch mean and variance of xt toward zero. We address this effect using sequencewise normalization of the inputs as proposed by Laurent et al. (2016); Amodei et al. (2015). That is, we share statistics over time for normalization 0 2000 4000 6000 8000 10000 12000 14000 16000 training steps1.41.61.82.02.22.4bits per characterLSTM BN-LSTM (a) Performance in bits-per-character on length100 subsequences of the Penn Treebank validation sequence during training. 100 200 300 400 500 600 700 800 900 1000 sequence length1.321.341.361.381.401.421.441.46mean bits per characterLSTM BN-LSTM, population statistics BN-LSTM, batch statistics(b) Generalization to longer subsequences of Penn Treebank using population statistics. The subsequences are taken from the test sequence. Figure 3: Penn Treebank evaluation 7 Published as a conference paper at ICLR 2017 0 100 200 300 400 500 600 700 800 training steps (thousands)0.00.20.40.60.81.0error rateLSTM train BN-LSTM train BN-everywhere train BN-e* train

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

24

BN-LSTM train BN-everywhere train BN-e* train BN-e** train LSTM valid BN-LSTM valid BN-everywhere valid BN-e* valid BN-e** valid (a) Error rate on the validation set for the Attentive Reader models on a variant of the CNN QA task (Hermann et al., 2015). As detailed in Appendix C, the theoretical lower bound on the error rate on this task is 43%. 0 50 100 150 200 250 300 350 400 training steps (thousands)0.10.20.30.40.50.60.70.80.91.0error rateLSTM train BN-e** train LSTM valid BN-e** valid(b) Error rate on the validation set on the full CNN QA task from Hermann et al. (2015). Figure 4: Training curves on the CNN question-answering tasks. of the input terms Wxxt, but notfor the recurrent terms Whhtor the cell output ct. Doing so avoids many issues involving degenerate statistics due to input sequence padding. Our fourth and final variant BN-e** is like BN-e* but bidirectional. The main difficulty in adapting

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

25

Our fourth and final variant BN-e** is like BN-e* but bidirectional. The main difficulty in adapting to bidirectional models also involves padding. Padding poses no problem as long as it is properly ignored (by not updating the hidden states based on padded regions of the input). However to perform the reverse application of a bidirectional model, it is common to simply reverse the padded sequences, thus moving the padding to the front. This causes similar problems as were observed on the sequential MNIST task (Section 5.1): the hidden states will not diverge during the initial timesteps and hence their variance will be severely underestimated. To get around this, we reverse only the unpadded portion of the input sequences and leave the padding in place. See Appendix C for hyperparameters and task details. Figure 4(a) shows the learning curves for the different variants of the attentive reader. BN-LSTM trains dramatically faster than the LSTM baseline. BN-everywhere in turn shows a significant improvement over BN-LSTM. In addition, both BN-LSTM and BN-everywhere show a generalization

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

26

benefit over the baseline. The validation curves have minima of 50.3%, 49.5% and 50.0% for the baseline, BN-LSTM and BN-everywhere respectively. We emphasize that these results were obtained without any tweaking – all we did was to introduce batch normalization. BN-e* and BN-e** converge faster yet, and reach lower minima: 47.1% and 43.9% respectively. Model CNN valid CNN test Attentive Reader (Hermann et al., 2015) 38.4 37.0 LSTM (ours) 45.5 45.0 BN-e** (ours) 37.9 36.3 Table 4: Error rates on the CNN question-answering task Hermann et al. (2015). We train and evaluate our best model, BN-e**, on the full task from (Hermann et al., 2015). On this dataset we had to reduce the number of hidden units to 120 to avoid severe overfitting. Training curves for BN-e** and a vanilla LSTM are shown in Figure 4(b). Table 4 reports performances of the early-stopped models. 8 Published as a conference paper at ICLR 2017 6 C ONCLUSION

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

27

the early-stopped models. 8 Published as a conference paper at ICLR 2017 6 C ONCLUSION Contrary to previous findings by Laurent et al. (2016); Amodei et al. (2015), we have demonstrated that batch-normalizing the hidden states of recurrent neural networks greatly improves optimization. Indeed, doing so yields benefits similar to those of batch normalization in feed-forward neural networks: our proposed BN-LSTM trains faster and generalizes better on a variety of tasks including language modeling and question-answering. We have argued that proper initialization of the batch normalization parameters is crucial, and suggest that previous difficulties (Laurent et al., 2016; Amodei et al., 2015) were due in large part to improper initialization. Finally, we have shown our model to apply to complex settings involving variable-length data, bidirectionality and highly nonlinear attention mechanisms. ACKNOWLEDGEMENTS The authors would like to acknowledge the following agencies for research funding and computing support: the Nuance Foundation, Samsung, NSERC, Calcul Québec, Compute Canada, the Canada Research Chairs and CIFAR. Experiments were carried out using the Theano (Team et al., 2016) and

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

28

Research Chairs and CIFAR. Experiments were carried out using the Theano (Team et al., 2016) and the Blocks and Fuel (van Merriënboer et al., 2015) libraries for scientific computing. We thank David Krueger, Saizheng Zhang, Ishmael Belghazi and Yoshua Bengio for discussions and suggestions. REFERENCES D. Amodei et al. Deep speech 2: End-to-end speech recognition in english and mandarin. arXiv:1512.02595 , 2015. M. Arjovsky, A. Shah, and Y . Bengio. Unitary evolution recurrent neural networks. arXiv:1511.06464 , 2015. Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv:1607.06450 , 2016. D. Bahdanau, K. Cho, and Y . Bengio. Neural machine translation by jointly learning to align and translate. ICLR , 2015. Y . Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is difficult. Neural Networks, IEEE Transactions on , 1994.

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

29

difficult. Neural Networks, IEEE Transactions on , 1994. K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv:1406.1078 , 2014. Junyoung Chung, Sungjin Ahn, and Yoshua Bengio. Hierarchical multiscale recurrent neural networks. arXiv:1609.01704 , 2016. A. Graves. Generating sequences with recurrent neural networks. arXiv:1308.0850 , 2013. David Ha, Andrew Dai, and Quoc V Le. Hypernetworks. arXiv:1609.09106 , 2016. K. M. Hermann, T. Kocisky, E. Grefenstette, L. Espeholt, W. Kay, M. Suleyman, and P. Blunsom. Teaching machines to read and comprehend. In NIPS , 2015. S. Hochreiter. Untersuchungen zu dynamischen neuronalen netzen. Master’s thesis , 1991.

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

30

S. Hochreiter. Untersuchungen zu dynamischen neuronalen netzen. Master’s thesis , 1991. S. Hochreiter and J Schmidhuber. Long short-term memory. Neural computation , 1997. Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. abs/1502.03167, 2015. D. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv:1412.6980 , 2014. D Krueger and R. Memisevic. Regularizing rnns by stabilizing activations. ICLR , 2016. 9 Published as a conference paper at ICLR 2017 David Krueger, Tegan Maharaj, János Kramár, Mohammad Pezeshki, Nicolas Ballas, Nan Rosemary Ke, Anirudh Goyal, Yoshua Bengio, Hugo Larochelle, and Aaron Courville. Zoneout: Regularizing rnns by randomly preserving hidden activations. arXiv:1606.01305 , 2016. C. Laurent, G. Pereyra, P. Brakel, Y . Zhang, and Y . Bengio. Batch normalized recurrent neural networks. ICASSP , 2016.

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

31

C. Laurent, G. Pereyra, P. Brakel, Y . Zhang, and Y . Bengio. Batch normalized recurrent neural networks. ICASSP , 2016. Quoc V Le, N. Jaitly, and G. Hinton. A simple way to initialize recurrent networks of rectified linear units. arXiv:1504.00941 , 2015. Qianli Liao and Tomaso Poggio. Bridging the gaps between residual learning, recurrent neural networks and visual cortex. arXiv:1604.03640 , 2016. M. Mahoney. Large text compression benchmark. 2009. M. P. Marcus, M. Marcinkiewicz, and B. Santorini. Building a large annotated corpus of english: The penn treebank. Comput. Linguist. , 1993. J. Martens and I. Sutskever. Learning recurrent neural networks with hessian-free optimization. In ICML , 2011. T. Mikolov, I. Sutskever, A. Deoras, H. Le, S. Kombrink, and J. Cernocky. Subword language modeling with neural networks. preprint , 2012. Yann Ollivier. Persistent contextual neural networks for learning symbolic data sequences. CoRR ,

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

32

modeling with neural networks. preprint , 2012. Yann Ollivier. Persistent contextual neural networks for learning symbolic data sequences. CoRR , abs/1306.0514, 2013. Marius Pachitariu and Maneesh Sahani. Regularization and nonlinearities for neural language models: when are they needed? arXiv:1301.5650 , 2013. Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. arXiv:1211.5063 , 2012. H. Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of statistical planning and inference , 2000. The Theano Development Team et al. Theano: A Python framework for fast computation of mathematical expressions. arXiv e-prints , abs/1605.02688, May 2016. T. Tieleman and G. Hinton. Lecture 6.5—RmsProp: Divide the gradient by a running average of its recent magnitude. COURSERA: Neural Networks for Machine Learning, 2012.

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

33

recent magnitude. COURSERA: Neural Networks for Machine Learning, 2012. Bart van Merriënboer, Dzmitry Bahdanau, Vincent Dumoulin, Dmitriy Serdyuk, David WardeFarley, Jan Chorowski, and Yoshua Bengio. Blocks and fuel: Frameworks for deep learning. CoRR , abs/1506.00619, 2015. URL http://arxiv.org/abs/1506.00619 . K. Xu, J. Ba, R. Kiros, A. Courville, R. Salakhutdinov, R. Zemel, and Y . Bengio. Show, attend and tell: Neural image caption generation with visual attention. arXiv:1502.03044 , 2015. L. Yao, A. Torabi, K. Cho, N. Ballas, C. Pal, H. Larochelle, and A. Courville. Describing videos by exploiting temporal structure. In ICCV , 2015. S. Zhang, Y . Wu, T. Che, Z. Lin, R. Memisevic, R. Salakhutdinov, and Y . Bengio. Architectural complexity measures of recurrent neural networks. arXiv:1602.08210 , 2016. 10 Published as a conference paper at ICLR 2017

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

34

complexity measures of recurrent neural networks. arXiv:1602.08210 , 2016. 10 Published as a conference paper at ICLR 2017 A C ONVERGENCE OF POPULATION STATISTICS 0.20 0.15 0.10 0.05 0.000.050.100.150.20mean of recurrent term 0.20 0.15 0.10 0.05 0.000.050.100.150.20mean of cell state 0 10 20 30 40 50 time steps0.00.51.01.52.0variance of recurrent term 0 10 20 30 40 50 time steps0.0010.0020.0030.0040.0050.0060.0070.0080.009variance of cell state Figure 5: Convergence of population statistics to stationary distributions on the Penn Treebank task. The horizontal axis denotes RNN time. Each curve corresponds to a single hidden unit. Only a random subset of units is shown. See Section 3 for discussion. B S ENSITIVITY TO INITIALIZATION OF In Section 4 we investigated the effect of initial on gradient flow. To show the practical implications of this, we performed several experiments on the pMNIST and Penn Treebank benchmarks. The resulting performances are shown in Figure 6.

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

35

The resulting performances are shown in Figure 6. ThepMNIST training curves confirm that higher initial values of are detrimental to the optimization of the model. For the Penn Treebank task however, the effect is gone. We believe this is explained by the difference in the nature of the two tasks. For pMNIST, the model absorbs the input sequence and only at the end of the sequence does it make a prediction on which it receives feedback. Learning from this feedback requires propagating the gradient all the way back through the sequence. In the Penn Treebank task on the other hand, the model makes a prediction at each timestep. At each step of the backward pass, a fresh learning signal is added to the backpropagated gradient. Essentially, the model is able to get off the ground by picking up short-term dependencies. This fails onpMNIST wich is dominated by long-term dependencies (Arjovsky et al., 2015). C T EACHING MACHINES TO READ AND COMPREHEND : TASK SETUP We evaluate the models on the question answering task using the CNN corpus (Hermann et al., 2015), with placeholders for the named entities. We follow a similar preprocessing pipeline as Hermann et al. (2015). During training, we randomly sample the examples with replacement and shuffle

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

36

the order of the placeholders in each text inside the minibatch. We use a vocabulary of 65829 words. We deviate from Hermann et al. (2015) in order to save computation: we use only the 4 most relevant sentences from the description, as identified by a string matching procedure. Both the training and validation sets are preprocessed in this way. Due to imprecision this heuristic sometimes strips the 11 Published as a conference paper at ICLR 2017 0 10000 20000 30000 40000 50000 training steps0.00.51.01.52.02.5cross entropyPermuted MNIST train gamma 0.10 gamma 0.30 gamma 0.50 gamma 0.70 gamma 1.00 0 10000 20000 30000 40000 50000 training steps0.00.51.01.52.02.5cross entropyPermuted MNIST valid gamma 0.10 gamma 0.30 gamma 0.50 gamma 0.70 gamma 1.00 0 5000 10000 15000 training steps0.800.850.900.951.001.051.10bits per characterPTB train gamma 0.10 gamma 0.30 gamma 0.50 gamma 0.70

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

37

gamma 0.10 gamma 0.30 gamma 0.50 gamma 0.70 gamma 1.00 0 5000 10000 15000 training steps1.001.021.041.061.081.10bits per characterPTB valid gamma 0.10 gamma 0.30 gamma 0.50 gamma 0.70 gamma 1.00 Figure 6: Training curves on pMNIST and Penn Treebank for various initializations of . answers from the passage, putting an upper bound of 57% on the validation accuracy that can be achieved. For the reported performances, the first three models (LSTM, BN-LSTM and BN-everywhere) are trained using the exact same hyperparameters, which were chosen because they work well for the baseline. The hidden state is composed of 240 units. We use stochastic gradient descent on minibatches of size 64, with gradient clipping at 10 and step rule determined by Adam (Kingma & Ba, 2014) with learning rate 8105. For BN-e* and BN-e**, we use the same hyperparameters except that we reduce the learning rate to 8104and the minibatch size to 40.

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

38

8104and the minibatch size to 40. D H YPERPARAMETER SEARCHES Table 5 reports hyperparameter values that were tried in the experiments. (a) MNIST and pMNIST Learning rate: 1e-2, 1e-3, 1e-4 RMSProp momentum: 0.5, 0.9 Hidden state size: 100, 200, 400 Initial : 1e-1, 3e-1, 5e-1, 7e-1, 1.0(b) Penn Treebank Learning rate: 1e-1, 1e-2, 2e-2, 1e-3 Hidden state size: 800, 1000, 1200, 1500, 2000 Batch size: 32, 64, 100, 128 Initial : 1e-1, 3e-1, 5e-1, 7e-1, 1.0 (c) Text8 Learning rate: 1e-1, 1e-2, 1e-3 Hidden state size: 500, 1000, 2000, 4000(d) Attentive Reader Learning rate: 8e-3, 8e-4, 8e-5, 8e-6 Hidden state size: 60, 120, 240, 280 Table 5: Hyperparameter values that have been explored in the experiments.

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1603.09025

39

Hidden state size: 60, 120, 240, 280 Table 5: Hyperparameter values that have been explored in the experiments. For MNIST and pMNIST, the hyperparameters were varied independently. For Penn Treebank, we performed a full grid search on learning rate and hidden state size, and later performed a sensitivity 12 Published as a conference paper at ICLR 2017 analysis on the batch size and initial . For the text8 task and the experiments with the Attentive Reader, we carried out a grid search on the learning rate and hidden state size. The same values were tried for both the baseline and our BN-LSTM. In each case, our reported results are those of the model with the best validation performance. 13

1603.09025

Recurrent Batch Normalization

We propose a reparameterization of LSTM that brings the benefits of batch normalization to recurrent neural networks. Whereas previous works only apply batch normalization to the input-to-hidden transformation of RNNs, we demonstrate that it is both possible and beneficial to batch-normalize the hidden-to-hidden transition, thereby reducing internal covariate shift between time steps. We evaluate our proposal on various sequential problems such as sequence classification, language modeling and question answering. Our empirical results show that our batch-normalized LSTM consistently leads to faster convergence and improved generalization.

http://arxiv.org/pdf/1603.09025

cs.LG

20160330

20170228

1605.07681

0

Convolutional Random Walk Networks for Semantic Image Segmentation Gedas Bertasius1, Lorenzo Torresani2, Stella X. Yu3, Jianbo Shi1 1University of Pennsylvania,2Dartmouth College,3UC Berkeley ICSI {gberta,jshi}@seas.upenn.edu lt@dartmouth.edu stella.yu@berkeley.edu Abstract Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by postprocessing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

1

issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard backpropagation. Our implementation of RWN requires just 131 additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling. 1. Introduction Fully convolutional networks (FCNs) were first introduced in [20] where they were shown to yield significant improvements in semantic image segmentation. Adopting the FCN approach, many subsequent methods have achieved even better performance [6, 27, 8, 12, 17, 6, 27, 8, 19, 23]. However, traditional FCN-based methods tend to suffer from several limitations. Large receptive fields in the Input DeepLab [6] DeepLab-CRF [6]

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

2

suffer from several limitations. Large receptive fields in the Input DeepLab [6] DeepLab-CRF [6] Figure 1: Examples illustrating shortcomings of prior semantic segmentation methods. Segments produced by FCNs are poorly localized around object boundaries, while Dense-CRF produce spatially-disjoint object segments. convolutional layers and the presence of pooling layers lead to low spatial resolution in the deepest FCN layers. As a result, their predicted segments tend to be blobby and lack fine object boundary details. We report in Fig. 1 some examples illustrating typical poor localization of objects in the outputs of FCNs. Recently, Chen at al. [6] addressed this issue by applying a Dense-CRF post-processing step [15] on top of coarse FCN segmentations. However, such approaches often fail to accurately capture semantic relationships between objects and lead to spatially fragmented segmentations (see an example in the last column of Fig. 1). To address these problems several recent methods integrated CRFs or MRFs directly into the FCN framework [27, 19, 17, 5]. However, these new models typically involve (1) a large number of parameters, (2) complex loss functions requiring specialized model training or (3) recurrent layers, which make training and testing more

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

3

(3) recurrent layers, which make training and testing more complicated. We summarize the most prominent of these approaches and their model complexities in Table 1. We note that we do not claim that using complex loss functions always makes the model overly complex and too difficult to use. If a complex loss is integrated into an FCN framework such that the FCN can still be trained in a standard fashion, and produce better results than using standard losses, then such a model is beneficial. However, in the context of prior segmentation methods [27, 19, 17, 5], such complex losses often require: 1) modifying the network structure (casting CNN into an RNN) [27, 5], or 2) using a complicated multi-stage learning scheme, where different layers are optimized during a different training stage [19, 17]. Due to such complex training procedures, 1arXiv:1605.07681v3 [cs.CV] 8 May 2017 [6] [5] [23] [27] [19] RWN requires post-processing? 3777 7 7 uses complex loss? 7773 3 7 requires recurrent layers? 7373 7 7 model size (in MB) 79 79 961 514 >1000 79

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

4

uses complex loss? 7773 3 7 requires recurrent layers? 7373 7 7 model size (in MB) 79 79 961 514 >1000 79 Table 1: Summary of recent semantic segmentation models. For each model, we report whether it requires: (1) CRF post-processing, (2) complex loss functions, or (3) recurrent layers. We also list the size of the model (using the size of Caffe [13] models in MB). We note that unlike prior methods, our RWN does not need post-processing, it is implemented using standard layers and loss functions, and it also has a compact model. which are adapted for specific tasks and datasets, these models can be quite difficult to adapt for new tasks and datasets, which is disadvantageous. Inspired by random walk methods [21, 3, 24], in this work we introduce a simple, yet effective alternative to traditional FCNs: a Convolutional Random Walk Network (RWN) that combines the strengths of FCNs and random walk methods. Our model addresses the issues of (1) poor localization around the boundaries suffered by FCNs and (2) spatially disjoint segments produced by dense CRFs. Additionally, unlike recent semantic segmentation

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

5

(1) poor localization around the boundaries suffered by FCNs and (2) spatially disjoint segments produced by dense CRFs. Additionally, unlike recent semantic segmentation approaches [23, 27, 19, 17], our RWN does so without significantly increasing the complexity of the model. Our proposed RWN jointly optimizes (1) pixelwise affinity and (2) semantic segmentation learning objectives that are linked via a novel random walk layer, which enforces spatial consistency in the deepest layers of the network. The random walk layer is implemented via matrix multiplication. As a result, RWN seamlessly integrates both affinity and segmentation branches, and can be jointly trained endto-end via standard back-propagation with minimal modifications to the existing FCN framework. Additionally, our implementation of RWN requires only 131additional parameters. Thus, the effective complexity of our model is the same as the complexity of traditional FCNs (see Table 1). We compare our approach to several variants of the DeepLab semantic segmentation system [6, 7], and show that our proposed RWN consistently produces better performance over these baselines for the tasks of semantic segmentation and scene labeling. 2. Related Work The recent introduction of fully convolutional networks

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

6

that our proposed RWN consistently produces better performance over these baselines for the tasks of semantic segmentation and scene labeling. 2. Related Work The recent introduction of fully convolutional networks (FCNs) [20] has led to remarkable advances in semantic segmentation. However, due to the large receptive fields and many pooling layers, segments predicted by FCNs tend to be blobby and lack fine object boundary details. Recently there have been several attempts to address these problems. These approaches can be divided into several groups.The work in [4, 6, 25, 14, 4] used FCN predictions as unary potentials in a separate globalization model that refines the segments using similarity cues based on regions or boundaries. One disadvantage of these methods is that the learning of the unary potentials and the training of the globalization model are completely disjoint. As a result, these methods often fail to capture semantic relationships between objects, which produces segmentation results that are spatially disjoint (see the right side of Fig. 1). To address these issues, several recent methods [27, 19, 17] have proposed to integrate a CRF or a MRF into the network, thus enabling end-to-end training of the joint model.

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

7

17] have proposed to integrate a CRF or a MRF into the network, thus enabling end-to-end training of the joint model. However, the merging of these two models leads to a dramatic increase in complexity and number of parameters. For instance, the method in [27], requires to cast the original FCN into a Recurrent Neural Network (RNN), which renders the model much bigger in size (see Table 1). A recent method [5] jointly predicts boundaries and segmentations, and then combines them using a recurrent layer, which also requires complex modifications to the existing FCN framework. The work in [19] proposes to use local convolutional layers, which leads to a significantly larger number of parameters. Similarly, the method in [17] proposes to model unary and pairwise potentials by separate multi-scale branches. This leads to a network with at least twice as many parameters as the traditional FCN and a much more complex multistage training procedure. In addition to the above methods, it is worth mentioning deconvolutional networks [23, 12], which use deconvolution and unpooling layers to recover fine object details from the coarse FCN predictions. However, in order to effectively recover fine details one must employ nearly as many

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

8

the coarse FCN predictions. However, in order to effectively recover fine details one must employ nearly as many deconvolutional layers as the number of convolutional layers, which yields a large growth in number of parameters (see Table 1). Unlike these prior methods, our implementation of RWN needs only 131additional parameters over the base FCN. These additional parameters represent only 0:0008% of the total number of parameters in the network. In addition, our RWN uses standard convolution and matrix multiplication. Thus, it does not need to incorporate complicated loss functions or new complex layers [27, 5, 19, 17]. Finally, unlike the methods in [4, 6, 25, 14] that predict and refine the segmentations disjointly, our RWN model jointly optimizes pixel affinity and semantic segmentation in an end-to-end fashion. Our experiments show that this leads to spatially smoother segmentations. 3. Background Random Graph Walks. Random walks are one of the most widely known and used methods in graph theory [21]. Most notably, the concept of random walks led to FC8 ActivationsDeepLab FCN Architecture RGB ImageSparse Pairwise L1 Distances Sparse Pixel Similarity Matrix1 x 1 x k conv Ground Truth Pixel Similarities

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

9

RGB ImageSparse Pairwise L1 Distances Sparse Pixel Similarity Matrix1 x 1 x k conv Ground Truth Pixel Similarities Exp Ground Truth Segmentation Segmentation Random Walk Layerˆy=AfAffinity Learning BranchPairwise L1 Distances n⇥n⇥mRow-Normalization:n⇥n⇥3 f2Rn2⇥mA2Rn2⇥n2W2Rn2⇥n2F2Rn2⇥n2⇥kFigure 2: The architecture of our proposed Random Walk Network (RWN) (best viewed in color). Our RWN consists of two branches: (1) one branch devoted to the segmentation predictions , and (2) another branch predicting pixel-level affinities. These two branches are then merged via a novel random walk layer that encourages spatially smooth segmentation predictions. The entire RWN is jointly optimized end-to-end via a standard back-propagation algorithm. the development of PageRank [24] and Personalized PageRank [3], which are widely used for many applications. Let G= (V;E)denote an undirected graph with a set of verticesVand a set of edges E. Then a random walk in

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

10

G= (V;E)denote an undirected graph with a set of verticesVand a set of edges E. Then a random walk in such graph can be characterized by the transition probabilities between its vertices. Let Wbe a symmetric nn affinity matrix, where ndenotes the number of nodes in the graph and where Wij2[0;1]denotes how similar the nodesiandjare. In the context of a semantic segmentation problem, each pixel in the image can be viewed as a separate node in the graph, where the similarity between two nodes can be evaluated according to some metric (e.g. color or texture similarity etc). Then let Dindicate a diagonalnnmatrix, which stores the degree values for each node:Dii=Pn j=1Wijfor alljexcepti=j. Then, we can express our random walk transition matrix as A=D1W. Given this setup, we want to model how the information in the graph spreads if we start at a particular node, and perform a random walk in this graph. Let ytbe an1vector denoting a node distribution at time t. In the context of the PageRank algorithm, ytmay indicate the rank estimates associated with each of the nWeb pages at time t. Then,

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

11

the PageRank algorithm, ytmay indicate the rank estimates associated with each of the nWeb pages at time t. Then, according to the random walk theory, we can spread the rank information in the graph by performing a one-step random walk. This process can be expressed as yt+1=Ayt, whereyt+1denotes a newly obtained rank distribution after one random walk step, the matrix Acontains the random walk transition probabilities, and ytis the rank distribution at time step t. Thus, we can observe that the information among the nodes can be diffused, by simply multiplying the random walk transition probability matrix A, with the rank distribution ytat a particular time t. This process can be repeated multiple times, until convergence is reached. For a more detailed survey please see [21, 24]. Difference from MRF/CRF Approaches. CRFs andMRFs have been widely used in structured prediction problems [16]. Recently, CRFs and MRFs have also been integrated into the fully convolutional network framework for semantic segmentation [27, 19, 17]. We want to stress that while the goals of CRF/MRF and random walk methods are the same (i.e. to globally propagate information in the graph structures), the mechanism to achieve this objective is very

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508

1605.07681

12

the same (i.e. to globally propagate information in the graph structures), the mechanism to achieve this objective is very different in these two approaches. While MRFs and CRFs typically employ graphs with a fixed grid structure (e.g., one where each node is connected to its four closest neighbors), random walk methods are much more flexible and can implement any arbitrary graph structure via the affinity matrix specification. Thus, since our proposed RWN is based on random walks, it can employ any arbitrary graph structure, which can be beneficial as different problems may require different graph structures. Additionally, to globally propagate information among the nodes, MRFs and CRFs need to employ approximate inference techniques, because exact inference tends to be intractable in graphs with a grid structure. Integrating such approximate inference techniques into the steps of FCN training and prediction can be challenging and may require lots of domain-specific modifications. In comparison, random walk methods globally propagate information among the nodes via a simple matrix multiplication. Not only is the matrix multiplication efficient and exact, but it is also easy to integrate into the traditional FCN framework for both training and prediction schemes. Additionally, due to

1605.07681

Convolutional Random Walk Networks for Semantic Image Segmentation

Most current semantic segmentation methods rely on fully convolutional networks (FCNs). However, their use of large receptive fields and many pooling layers cause low spatial resolution inside the deep layers. This leads to predictions with poor localization around the boundaries. Prior work has attempted to address this issue by post-processing predictions with CRFs or MRFs. But such models often fail to capture semantic relationships between objects, which causes spatially disjoint predictions. To overcome these problems, recent methods integrated CRFs or MRFs into an FCN framework. The downside of these new models is that they have much higher complexity than traditional FCNs, which renders training and testing more challenging. In this work we introduce a simple, yet effective Convolutional Random Walk Network (RWN) that addresses the issues of poor boundary localization and spatially fragmented predictions with very little increase in model complexity. Our proposed RWN jointly optimizes the objectives of pixelwise affinity and semantic segmentation. It combines these two objectives via a novel random walk layer that enforces consistent spatial grouping in the deep layers of the network. Our RWN is implemented using standard convolution and matrix multiplication. This allows an easy integration into existing FCN frameworks and it enables end-to-end training of the whole network via standard back-propagation. Our implementation of RWN requires just $131$ additional parameters compared to the traditional FCNs, and yet it consistently produces an improvement over the FCNs on semantic segmentation and scene labeling.

http://arxiv.org/pdf/1605.07681

cs.CV

20160524

20170508