Learning Fast Algorithms for Linear Transforms Using Butterfly Factorizations

TL;DR

This paper introduces a method to automatically learn fast algorithms for structured linear transforms, such as Fourier transforms, using butterfly factorizations, achieving near-optimal efficiency and improved performance in machine learning tasks.

Contribution

It presents a parameterization that can automatically discover efficient divide-and-conquer algorithms for structured transforms, including the FFT, without manual hand-crafting.

Findings

Recovers the O(N log N) FFT algorithm to machine precision

Achieves 4X faster inference speed in neural network compression

Reduces parameters by 40X compared to unstructured matrices

Abstract

Fast linear transforms are ubiquitous in machine learning, including the discrete Fourier transform, discrete cosine transform, and other structured transformations such as convolutions. All of these transforms can be represented by dense matrix-vector multiplication, yet each has a specialized and highly efficient (subquadratic) algorithm. We ask to what extent hand-crafting these algorithms and implementations is necessary, what structural priors they encode, and how much knowledge is required to automatically learn a fast algorithm for a provided structured transform. Motivated by a characterization of fast matrix-vector multiplication as products of sparse matrices, we introduce a parameterization of divide-and-conquer methods that is capable of representing a large class of transforms. This generic formulation can automatically learn an efficient algorithm for many important transforms; for example, it recovers the $O(N \log N)$ Cooley-Tukey FFT algorithm to machine precision, for dimensions $N$ up to $1024$. Furthermore, our method can be incorporated as a lightweight replacement of generic matrices in machine learning pipelines to learn efficient and compressible transformations. On a standard task of compressing a single hidden-layer network, our method exceeds the classification accuracy of unconstrained matrices on CIFAR-10 by 3.9 points -- the first time a structured approach has done so -- with 4X faster inference speed and 40X fewer parameters.