Abstract

This article presents a generic approach to convolution that significantlydiffers from conventional methodologies in the current Machine Learningliterature. The approach, in its mathematical aspects, proved to be clear andconcise, particularly when high-order tensors are involved. In this context, arational theory of regression in neural networks is developed, as a frameworkfor a generic view of sparse convolutional neural networks, the primary focusof this study. As a direct outcome, the classic Backpropagation Algorithmic isredefined to align with this rational tensor-based approach and presented inits simplest, most generic form.

 

Quick Read (beta)