John C
commited on
Commit
•
037e8af
1
Parent(s):
f9dbc26
Update README.md
Browse files
README.md
CHANGED
@@ -7,7 +7,41 @@ language:
|
|
7 |
pipeline_tag: audio-classification
|
8 |
---
|
9 |
|
|
|
10 |
|
11 |
-
|
|
|
12 |
|
13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
7 |
pipeline_tag: audio-classification
|
8 |
---
|
9 |
|
10 |
+
# SNN Voice
|
11 |
|
12 |
+
SNN Voice is the repository for my paper: "Retraining SNN Conversions: CNN to SNN for Audio Classification Tasks".
|
13 |
+
The models are deposited here for archival.
|
14 |
|
15 |
+
See our code repository for implementation @ GitHub: https://github.com/Eve-ning/snn_voice
|
16 |
+
|
17 |
+
## Abstract
|
18 |
+
|
19 |
+
Efficient yet powerful models are in high demand for its portability and affordability.
|
20 |
+
Amongst other methods such as model-pruning, is limiting neural network operations to
|
21 |
+
sparse event-driven spikes: Spiking Neural Networks (SNNs) aims to unravel a new di-
|
22 |
+
rection in machine learning research. A significant amount of SNN literature straddles
|
23 |
+
upon mature works of artificial neural networks (ANNs) by migrating its architecture
|
24 |
+
and parameters into SNNs, optimizing the migration to retain as much performance as
|
25 |
+
possible. We spearhead a novel approach: the architecture is migrated and retrained
|
26 |
+
from scratch. We hypothesize that this new direction will unravel concepts that cur-
|
27 |
+
rently bottlenecks improvements in the field of SNN conversions. Furthermore, alike
|
28 |
+
Transfer Learning, inspire future efforts of fine-tuning a well converted model through
|
29 |
+
training.
|
30 |
+
|
31 |
+
This paper presents our analysis of training converted Convolutional Neural Networks
|
32 |
+
(CNNs) to SNNs on audio classification models. Results show that (1) SNN conver-
|
33 |
+
sions consistently underperforms CNNs marginally during training, however we also
|
34 |
+
show that model complexity has a possible association with this margin. (2) SNN con-
|
35 |
+
verts doesn’t necessarily approach the performance of its CNN counterparts asymptot-
|
36 |
+
ically by increasing the number of time-steps. (3) SNN training from scratch is costly
|
37 |
+
and impractical with current hardware and dedicated SNN optimization techniques are
|
38 |
+
necessary. (4) Enabling the SNN membrane decay rate to be learned doesn’t signifi-
|
39 |
+
cantly affect performance. This paper provides valuable insights into the perspective of
|
40 |
+
retraining converted SNNs for audio classification, and serves as a reference for future
|
41 |
+
studies and hardware implementation benchmarks.
|
42 |
+
|
43 |
+
## SNN Voice Trained Models
|
44 |
+
|
45 |
+
**Important**
|
46 |
+
|
47 |
+
We only upload Piczak's SNN Model for 15 time-steps with Learnable Beta = False due to it's extremely large size (300MB)
|