Deep Learning Compression & Hugging Face Hub

Deep Neural Networks are powerful machine learning models that have obtained excellent performance for numerous applications. However, they may require a large amount of computational and memory resources leading to a high energy consumption making their use for embedded systems difficult. In this talk, I will present two different approaches that we proposed recently to reduce this complexity. This first is a post-training quantization method that approximates the filters of a CNN using dyadic rationals such that all multiplications of the model can be replaced by bit shifts and additions. The second approach is a compression-aware training algorithm, that performs filter pruning on a CNN by introducing a new sparsity-inducing regularization term based on the l1/l2 pseudo norm. Both approach have been applied to standard computer vision benchmarks and can considerably reduce the complexity of CNN with only a small precision loss.