Fast Training of Convolutional Networks through FFTs

TL;DR

This paper introduces a GPU-accelerated algorithm that significantly speeds up convolutional network training and inference by leveraging Fourier transforms to compute convolutions more efficiently, reducing computational costs substantially.

Contribution

The authors propose a novel method that computes convolutions in the Fourier domain, enabling faster training and inference of convolutional networks on GPUs, with over an order of magnitude improvement.

Findings

Achieved over tenfold speedup in training and inference times.

Demonstrated effective GPU implementation of Fourier-based convolution.

Reduced computational costs for large-scale convolutional networks.

Abstract

Convolutional networks are one of the most widely employed architectures in computer vision and machine learning. In order to leverage their ability to learn complex functions, large amounts of data are required for training. Training a large convolutional network to produce state-of-the-art results can take weeks, even when using modern GPUs. Producing labels using a trained network can also be costly when dealing with web-scale datasets. In this work, we present a simple algorithm which accelerates training and inference by a significant factor, and can yield improvements of over an order of magnitude compared to existing state-of-the-art implementations. This is done by computing convolutions as pointwise products in the Fourier domain while reusing the same transformed feature map many times. The algorithm is implemented on a GPU architecture and addresses a number of related challenges.