Binary Convolution Model for Image Classification

Binary Neural Networks present an opportunity for developing Neural Networks that require less computing power as well as energy. This is done through the use of binary values for weights and inputs. This research presents an architecture for image classification where the image and the Convolutional Layers are binarized. The result of the binary convolution is then fed to a non-binary network. The latter network is meant to be the standard Fully Connected Layer that is included as the final stage of a typical Convolutional Neural Network (CNN), after which a Soft Max function does the final classification. The Network is trained using the MNIST data set. The resulting Network achieved an accuracy of more than 82% using validation data.