Recursive Autoconvolution for Unsupervised Learning of Convolutional Neural Networks

TL;DR

This paper introduces recursive autoconvolution, a physics-inspired operator, to enhance unsupervised learning of CNN filters, achieving state-of-the-art results on multiple image classification benchmarks.

Contribution

It proposes a novel recursive autoconvolution technique to improve filter learning in unsupervised CNNs and designs a network with over 600k features and shared filters for efficiency.

Findings

Achieved state-of-the-art results on MNIST, CIFAR-10, CIFAR-100, STL-10

Enhanced filter discriminability using recursive autoconvolution

Reduced parameters while extracting a large number of features

Abstract

In visual recognition tasks, such as image classification, unsupervised learning exploits cheap unlabeled data and can help to solve these tasks more efficiently. We show that the recursive autoconvolution operator, adopted from physics, boosts existing unsupervised methods by learning more discriminative filters. We take well established convolutional neural networks and train their filters layer-wise. In addition, based on previous works we design a network which extracts more than 600k features per sample, but with the total number of trainable parameters greatly reduced by introducing shared filters in higher layers. We evaluate our networks on the MNIST, CIFAR-10, CIFAR-100 and STL-10 image classification benchmarks and report several state of the art results among other unsupervised methods.