Memory Optimization of CNN for Reconfigurable Architectures

Abstract

Deep neural networks have become the most advanced method for solving computer vision problems. Although they are much Powerful, big amount of weights requires significant storage space. So, these deep Neural networks have high computational cost and require more memory resources which makes them hard to run it on systems with less hardware resources. As of Today, the basic memory optimization technique after quantization is Huffman encoding. However, the lossless encoding schemes based on Huffman encoding are significantly slower than others. So in our proposed design in order to reduce memory requirement and make decoding phase fast we use two stage process Quantization and a Proposed Encoding Scheme which works together to decrease the memory requirement of Neural Network. Our method first quantizes the layer weights and then encoding scheme is applied which further reduces the storage requirement. In our work for testing purpose we use CIFAR10 and MNIST datasets. However, in terms of compression ratio our encoding scheme is not as efficient as Huffman encoding but in terms of decoding time it is way more effective than Huffman encoding.