Trainable back-propagated functional transfer matrices

Abstract

Functional transfer matrices consist of real functions with trainable parameters. In this work, functional transfer matrices are used to model functional connections in neural networks. Different from linear connections in conventional weight matrices, the functional connections can represent nonlinear relations between two neighbouring layers. Neural networks with the functional connections, which are called functional transfer neural networks, can be trained via back-propagation. On the two spirals problem, the functional transfer neural networks are able to show considerably better performance than conventional multi-layer perceptrons. On the MNIST handwritten digit recognition task, the performance of the functional transfer neural networks is comparable to that of the conventional model. This study has demonstrated that the functional transfer matrices are able to perform better than the conventional weight matrices in specific cases, so that they can be alternatives of the conventional ones.