Functional Machine Learning with Hasktorch: Produce Functional Machine Learning Model Reference Implementations

<p>Neural networks are a powerful tool in machine learning. An integral part of any network is its architecture. However, structuring and comparing new architectures is non-trivial. We describe a toolkit for structuring neural networks that leverages a special class of Naperian functors and functions between them. The explored class admits a canonical traversable structure which allows common neural network constructs, and shows the literal correspondence between higher-order functions and some architectures. We also show our toolkit is compatible with monadic computations necessary for training and evaluation. Our methods allow for expressive, unified, and elegant constructions.</p>