Matching Convolutional Neural Networks without Priors about Data

TL;DR

This paper introduces an extended CNN framework capable of handling graph-structured data, achieving comparable accuracy on images and improved results on fMRI data without relying on data priors.

Contribution

It presents a novel extension of CNNs to graph data, including strided convolutions and data augmentation, without requiring prior knowledge of data structure.

Findings

Matches state-of-the-art CNN accuracy on images

Achieves significant accuracy gains on fMRI data

Extends CNN applicability to graph-structured data

Abstract

We propose an extension of Convolutional Neural Networks (CNNs) to graph-structured data, including strided convolutions and data augmentation on graphs. Our method matches the accuracy of state-of-the-art CNNs when applied on images, without any prior about their 2D regular structure. On fMRI data, we obtain a significant gain in accuracy compared with existing graph-based alternatives.