Title | ||
---|---|---|
2<sup>n</sup>+1-valued SSS-Net: Uniform Shift, Channel Sparseness, and Channel Shuffle |
Abstract | ||
---|---|---|
Convolutional neural networks (CNNs) are primarily a cascaded set of pattern recognition filters, which are trained by big data. It enables us to solve complex problems of computer vision applications. A conventional CNN requires numerous parameters (weights) and computations. In this study, we propose SSS-Net using uniform channel shift, weight sparseness, and channel shuffle operations. We prove that a conventional k × k kernel convolution can be divided into k
<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>
of channel shift operations and a point-wise (1 × 1) convolution. We develop a 2
<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">n</sup>
+1-valued quantization with zero weight sparseness technique. We investigate a weight distribution for a post-training CNN, and almost all weights are close to zero. We eliminate such small weights to reduce the model size and non-zero weights are efficiently quantized by 2
<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">n</sup>
-valued one. We show an algorithm for a uniform shift with quantization. Since a uniform shift operation requires no multiplications, the amount of computation becomes zero. We train SSS (Shift, Sparseness, and Shuffle)-Net by ImageNet 2012 benchmark image. Compared with existing CNN models, our SSS-Net is a smaller model size and MAC operations, while it has considerable recognition accuracy. |
Year | DOI | Venue |
---|---|---|
2020 | 10.1109/ISMVL49045.2020.000-5 | 2020 IEEE 50th International Symposium on Multiple-Valued Logic (ISMVL) |
Keywords | DocType | ISSN |
Deep Learning,CNN,Multi valued Logic | Conference | 0195-623X |
ISBN | Citations | PageRank |
978-1-7281-5407-7 | 0 | 0.34 |
References | Authors | |
0 | 1 |
Name | Order | Citations | PageRank |
---|---|---|---|
Hiroki Nakahara | 1 | 155 | 37.34 |