Fast Training of Light Binary Convolutional Neural Networks using Chainer and Cupy
Radu Dogaru, Ioana Dogaru · 2020
Light binary convolutional neural networks (LB-CNN) are particularly useful when implemented in hardware technologies, such as FPGA. In this paper, such a network and its implementation using the Chainer machine learning framework is presented. Up to three convolutional layers, each provided with binary convolution kernels, can be defined forming a nonlinear expander for the images to be categorized. Effective training is done in a single output linear layer (Adaline) using the very fast extreme learning machine training with GPU acceleration. It is shown that a significant speedup of more than 60 times can be achieved while employing the CHAINER environment instead the more traditional Keras/Tensorflow. Such fast training allows easy tuning of the randomly generated binary kernels in a loop. Results for widely used datasets including MNIST, GTRSB, ORL, VGG etc. are given, showing very good compromise between accuracy and complexity. Particularly, for face recognition problems the model provides up to 100% accuracies. Such LB-CNN models are an excellent solutions for low-power applications where bit-quantization is an important option to reduce implementation complexity and energy.