Can I read Accelerating CNN Training with FFTs on EtoBox?

Accelerating CNN Training with FFTs by Tim Mazumdar is a document available to read on EtoBox.

What is Accelerating CNN Training with FFTs about?

This document proposes a method to accelerate training and inference for convolutional neural networks using fast Fourier transforms (FFTs). The key idea is to compute convolutions between pairs of feature maps and kernels as pointwise products in the Fourier domain, reusing the Fourier transforms of the feature maps. This can provide over an order of magnitude speedup compared to direct convolution. The algorithm works by computing the FFTs of all input feature maps and kernels once, then performing all co

Author

Tim Mazumdar

Language

EN