ConvART: Improving Adaptive Resonance Theory for Unsupervised Image Clustering

Ilia Sucholutsky, Matthias Schonlau

Abstract: While supervised learning techniques have become increasingly adept at separating images into different classes, these techniques require large amounts of labelled data which may not always be available. We propose a novel method for unsupervised image clustering by combining Adaptive Resonance Theory (ART) with techniques from Convolutional Neural Networks (CNN). ART networks are unsupervised clustering algorithms that have high stability in preserving learned information while quickly learning new information. Meanwhile, a major property of CNNs is their translation and distortion invariance, which has led to their success in the domain of vision problems. By embedding convolutional layers into an ART network, the useful properties of both networks can be leveraged to identify different clusters within unlabelled image datasets and classify images into these clusters. In exploratory experiments, we demonstrate that this method greatly increases the performance of unsupervised ART networks on a benchmark image dataset.

Keywords: adaptive resonance theory, unsupervised learning, convolutional neural networks, image classification

TL;DR: We aim to simultaneously leverage the benefits of both CNN and ART by embedding convolutional layers into ART networks in order to create a novel unsupervised method for discovering classes in image datasets. We demonstrate that this method greatly increases the performance of unsupervised ART networks on a benchmark image dataset.