A Feature Map Lossless Compression Framework for Convolutional Neural Network Accelerators

This paper proposes a predictor-based lossless compression algorithm for the feature maps present within convolutional neural networks (CNNs), which provides the possibility to solve the system bandwidth bottleneck and excessive power consumption problem of hardware acceleration. It is also an algorithm-hardware co-design methodology, yielding a hardware-friendly compression approach with low power consumption. The performance of the algorithm is evaluated in the detection, recognition, and segment CNN tasks respectively. Results show that an average compression ratio of 3.03x and a gain of nearly 50% over existing methods can be achieved for VGG-16; 2.78x and a gain of around 51% for ResNet-18; 2.45 and a gain of nearly 38% for SegNet.