B-VGG16: Binary Quantized Convolutional Neuronal Network for image classification

Abstract

In this work, a Binary Quantized Convolution neural network for image classification is trained and evaluated. Binarized neural networks reduce the amount of memory, and it is possible to implement them with less hardware than those that use real value variables (Floating Point 32 bits). This type of network can be implemented in embedded systems, such as FPGA. A quantization-aware training was performed, to compensate for the errors caused by the loss of precision of the parameters. The model obtained an evaluation accuracy of 88% with the CIFAR-10 evaluation set.

Keywords

• redes neuronales de convolución
• clasificación
• cuantización