| Summary: | 碩士 === 國立交通大學 === 資訊科學與工程研究所 === 107 === Recently, convolution neural networks (CNN) has been widely applied on image classification and object detection. Generally, the convolution layer requires lots of arithmetic operations in whole CNN model. Thus, the convolution layer plays an significant role in CNN hardware. In this thesis, we propose a new CNN accelerator to reach the low area and decrease the operation in convolution layer. To the best of our knowledge, we adopt binary-weights and activation (BWA) method to bound the filter weighted value into binary value in forward propagation part. Based on the BWA method, we found that the multiplication operation can be substituted into XNOR gate operation to significantly lower the calculation complexity. Furthermore, XNOR gate operation not only achieve low space and time complexities but reaches efficient energy. Besides, we provide flexible multi-size filter, .i.e., 1 × 1, 2 × 2, …, 7 × 7. For the storage of input feature maps and filter, we adopt 32x32-based computation mechanism.
|
|---|
