2018/12/10 by Georgiadis Georgios, Georgiadis, Georgios · 1 citation
Computer Science · #Advanced Neural Network Applications #Advanced Image and Video Retrieval Techniques #Medical Image Segmentation Techniques
paper · pdf · doi:10.48550/arxiv.1812.04056
The deep learning revolution brought us an extensive array of neural network\narchitectures that achieve state-of-the-art performance in a wide variety of\nComputer Vision tasks including among others, classification, detection and\nsegmentation. In parallel, we have also been observing an unprecedented demand\nin computational and memory requirements, rendering the efficient use of neural\nnetworks in low-powered devices virtually unattainable. Towards this end, we\npropose a three-stage compression and acceleration pipeline that sparsifies,\nquantizes and entropy encodes activation maps of Convolutional Neural Networks.\nSparsification increases the representational power of activation maps leading\nto both acceleration of inference and higher model accuracy. Inception-V3 and\nMobileNet-V1 can be accelerated by as much as 1.6\× with an increase in\naccuracy of 0.38 % and 0.54 % on the ImageNet and CIFAR-10 datasets\nrespectively. Quantizing and entropy coding the sparser activation maps lead to\nhigher compression over the baseline, reducing the memory cost of the network\nexecution. Inception-V3 and MobileNet-V1 activation maps, quantized to 16\nbits, are compressed by as much as 6\× with an increase in accuracy of\n0.36 % and 0.55 % respectively.\n