Slimmer CNNs Through Feature Approximation and Kernel Size Reduction

Convolutional Neural Networks (CNNs) have been shown to achieve state of the art results on several image processing tasks such as classification, localization, and segmentation. fully connected layers form building blocks these networks. The convolution are responsible for majority computations even though they fewer parameters. As inference is used much more than training (which happens only once), it important reduce network this phase. This work presents a systematic procedure trim CNNs by identifying least features in replacing them either with approximations or kernels reduced size. We also propose an algorithm integrate lower kernel approximation technique given accuracy budget. show that using linear method can 15 – 80% savings median 52% reduction while 33 95% 65% required number marginal 1% loss across benchmark datasets. demonstrated proposed methods VGG-16 architecture various On we achieved 4.2 -45% MAC (with 18.5%) 0.5% accuracy. how existing hardware accelerator DNNs (DianNao) be modified low added complexity take advantage approximations, estimate speedups obtained way custom embedded hardware.