Compositional Deep Learning in Futhark at FHPNC 2019
18 August 2019
Our paper Compositional Deep Learning in Futhark, with Duc Minh Tran and Troels Henriksen, was accepted for presentation at FHPNC 2019 in Berlin on August 18.
We show how to build neural networks by composing typed library components, including dense and convolutional layers. The same structure supports training through backpropagation and prediction through forward propagation. Futhark removes the higher-order functions and module abstractions during compilation, leaving opportunities for fusion and other GPU optimisations. The result connects a modular way of describing networks with efficient generated code.