Authors:
- Meir Goldberg
- Almog Hadad
A Novel Approach for Neural Network Compression. Using ALDS compression to improve IMP compression Maintain accuracy.
Per layer compression:
-
Weight tensor is folded into a matrix
-
Uses an SVD to compress the weights
Global compression:
-
Ensure global compression rate across all layers
-
Maintain model accuracy
-
Demonstrated to achieve accurate compression of over 60% on known neural networks
- Based on the lottery ticket hypothesis: A dense randomly initialized feed-forward network contains subnetworks (known as winning tickets) that when trained in isolation reach a test accuracy comparable to the original network in a similar number of iterations.
Identify the winning ticket:
-
Train the network
-
Prune the smallest magnitude weights
-
Winning tickets can be less than 10-20% the size of the full network
-
6 convolutional layers
-
3 FC layers
-Using CIFAR-10 as our dataset
-
Run compression with a fixed value
-
Reset the net
-
Repeat
Real understanding of the results โ further analysis requires more resources Our assumption:
-
ALDS reduced the values of insignificant parameters
-
Maybe also increased the values of significant parameters
-
This assisted IMP in pruning the more insignificant parameters
-
also likely: ALDS prevented IMP from removing entire layers