This paper presents the efficient acceleration techniques for the binarized convolutional neural network (BCNN). First, multiple binarized data are stored n a packed type so as the operations between them can be implemented effectively by bit-wise ope...
This paper presents the efficient acceleration techniques for the binarized convolutional neural network (BCNN). First, multiple binarized data are stored n a packed type so as the operations between them can be implemented effectively by bit-wise operations for the packed types. Secondly, multiple processing steps involved in each layer of the BCNN are merged into a simple but mathematically-equivalent one. Finally, the convolution operations of the binarized data are accelerated by employing a dedicated hardware accelerator. The experimental results show that the processing time of the BCNN is reduced by 94.3% using the proposed techniques in comparison to the straightforward implementation using only software.