Continuous Convolutional Neural Networks for Image Classification

Vitor Guizilini, Fabio Ramos

Abstract: This paper introduces the concept of continuous convolution to neural networks and deep learning applications in general. Rather than directly using discretized information, input data is first projected into a high-dimensional Reproducing Kernel Hilbert Space (RKHS), where it can be modeled as a continuous function using a series of kernel bases. We then proceed to derive a closed-form solution to the continuous convolution operation between two arbitrary functions operating in different RKHS. Within this framework, convolutional filters also take the form of continuous functions, and the training procedure involves learning the RKHS to which each of these filters is projected, alongside their weight parameters. This results in much more expressive filters, that do not require spatial discretization and benefit from properties such as adaptive support and non-stationarity. Experiments on image classification are performed, using classical datasets, with results indicating that the proposed continuous convolutional neural network is able to achieve competitive accuracy rates with far fewer parameters and a faster convergence rate.

TL;DR: This paper proposes a novel convolutional layer that operates in a continuous Reproducing Kernel Hilbert Space.

Keywords: convolutional neural networks, image classification, deep learning, feature representation, hilbert maps, reproducing kernel hilbert space