The idea in plain language.
Remove unnecessary connections, share weight values, and encode the remaining information efficiently.
How it works
The pipeline first removes connections judged unnecessary, then quantizes the remaining weights so several connections can share a value. Retraining helps the compressed network recover useful behavior. A coding stage further reduces the storage needed to represent the model. The paper evaluates how these stages work together on image-classification networks.
What to keep in mind
A smaller stored model is not automatically faster on every processor. Runtime gains depend on whether the deployment system can exploit sparsity and compressed representations.
Source: Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.