Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -11,4 +11,55 @@ source_datasets:
|
|
11 |
- extended|mnist
|
12 |
annotations_creators:
|
13 |
- machine-generated
|
14 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
- extended|mnist
|
12 |
annotations_creators:
|
13 |
- machine-generated
|
14 |
+
---
|
15 |
+
|
16 |
+
|
17 |
+
## Mnist-Ambiguous
|
18 |
+
|
19 |
+
This dataset contains mnist-like images, but with an unclear ground truth. For each image, there are two classes which could be considered true.
|
20 |
+
Robust and uncertainty-aware DNNs should thus detect and flag these issues.
|
21 |
+
|
22 |
+
### Features
|
23 |
+
Same as mnist, the supervised dataset has an `image` (28x28 int array) and a `label` (int).
|
24 |
+
|
25 |
+
Additionally, the following features are exposed for your convenience:
|
26 |
+
|
27 |
+
- `text_label` (str): A textual representation of the probabilistic label, e.g. `p(Pullover)=0.54, p(Shirt)=0.46`
|
28 |
+
- `p_label` (list of floats): Ground-Truth probabilities for each class (two nonzero values for our ambiguous images)
|
29 |
+
- `is_ambiguous` (bool): Flag indicating if this is one of our ambiguous images (see 'splits' below)
|
30 |
+
|
31 |
+
### Splits
|
32 |
+
We provide four splits:
|
33 |
+
|
34 |
+
- `test`: 10'000 ambiguous images
|
35 |
+
- `train`: 10'000 ambiguous images - adding ambiguous images to the training set makes sure test-time ambiguous images are in-distribution.
|
36 |
+
- `test_mixed`: 20'000 images, consisting of the (shuffled) concatenation of our ambiguous `test` test and the nominal mnist test set by LeCun et. al.,
|
37 |
+
- `train_mixed`: 70'000 images, consisting
|
38 |
+
|
39 |
+
For research targeting explicitly aleatoric uncertainty, we recommend training the model using `train_mixed`.
|
40 |
+
Otherwise, our `test` set will lead to both epistemic and aleatoric uncertainty.
|
41 |
+
Note that in related literature, these 'mixed' splits are sometimes denoted as *dirty* splits.
|
42 |
+
|
43 |
+
### Assessment and Validity
|
44 |
+
For a brief discussion of the strength and weaknesses of this dataset,
|
45 |
+
including a quantitative comparison to the (only) other ambiguous datasets available in the literature, we refer to our paper.
|
46 |
+
|
47 |
+
### Paper
|
48 |
+
Pre-print here: [https://arxiv.org/abs/2207.10495](https://arxiv.org/abs/2207.10495)
|
49 |
+
|
50 |
+
Citation:
|
51 |
+
```
|
52 |
+
@misc{https://doi.org/10.48550/arxiv.2207.10495,
|
53 |
+
doi = {10.48550/ARXIV.2207.10495},
|
54 |
+
url = {https://arxiv.org/abs/2207.10495},
|
55 |
+
author = {Weiss, Michael and Gómez, André García and Tonella, Paolo},
|
56 |
+
title = {A Forgotten Danger in DNN Supervision Testing: Generating and Detecting True Ambiguity},
|
57 |
+
publisher = {arXiv},
|
58 |
+
year = {2022}
|
59 |
+
}
|
60 |
+
```
|
61 |
+
|
62 |
+
### License
|
63 |
+
As this is a derivative work of mnist, which is CC-BY-SA 3.0 licensed, our dataset is released using the same license.
|
64 |
+
|
65 |
+
|