We present a design pattern for composing deep learning networks in a typed, higher-order fashion. The exposed library functions are generically typed and the composition structure allows for networks to be trained (using back-propagation) and for trained networks to be used for predicting new results (using forward-propagation). Individual layers in a network can take different forms ranging over dense sigmoid layers to convolutional layers.

The paper discusses different typing techniques aimed at enforcing proper use and composition of networks.

The approach is implemented in Futhark, a data-parallel functional language and compiler targeting GPU architectures, and we demonstrate that Futhark’s elimination of higher-order functions and modules leads to efficient generated code.