TCB Convolution: Ternary-Coded Binarized Convolutions with Fixed-Point Filters

Winograd convolution has almost become a standard for reducing computation complexity for convolutional neural network. However, its efficiency is limited due to the surge in the number of additions. In this paper, we analyze the upper bound of addition count of Winograd convolution with fixed-point filters and an estimated arithmetic weight density, and then proposed a ternary-coded binarized (TCB) convolution that can eliminate all multiplications and reduce more than 99% of addition operations.