/NeuroPhi

Neuro Phi: Parallelization of convolutional neural networks by use of FFT and other optimizations.

Primary LanguageCuda

NeuroPhi

Final project for 15-418 Parallel Computer Architecture and Programming. NeuroPhi is a framework for training convolutional neural networks on the Xeon Phi. We made use of the Fast Fourier Transform in performing the convolution and proposed several ways to improve memory efficiency. Also implemented parallel batch training and other tricks. Achieved state of the art training speeds at time of writing.