Leveraging Filter Correlations for Deep Model Compression

This Blog is to explain Leveraging Filter Correlations for Deep Model Compression Paper accepted in WACV 2020.

Approach

Iteratively identify pairs of filters with the largest pairwise correlations.
The model is re-optimized to make the filters in these pairs maximally correlated.
After discarding the filters in each round, we further fine tune the model (for smaller number of epochs) to recover from the potential small loss incurred due to the compression.

Episode Selection

N- pairs of filters are selected from each layer with maximum correlation.
Correlation between filters:

This selected set of filter pairs from all the layers is called one episode.

Optimization

We optimize this cost function with the new regularizer applied to the selected episode.
Our new regularizer (CSt) is as follows:

This new regularizer is added to the original cost function so our new objective function is as follows:

Pruning and Fine-Tuning

After increasing the correlation between filters in each pair we can prune one filter from each pair.
Our model has a reduced set of the parameter Θ’.

Where pi is the set of filters finally selected to be removed from the model.
Further, we fine-tune the model w.r.t. the parameter Θ’.

References

[1] .Singh, Pravendra, et al. “Leveraging filter correlations for deep model compression.” The IEEE Winter Conference on Applications of Computer Vision. 2020.