deno-ann

A simple artifical neural network library I wrote to study deep learning.

This is a library of routines written in Typescript for buliding and training artificial neural networks, loosely based on the code samples in Deep Learning from Scratch by Seth Weidman.

The neural network routines here are not based on automatic differentiation. The codes to propagate gradients backward have to be written manually. Writing backward pass manually is a tidious process.

Also only the simplest models can be trained in a reasonable time because the numerical routines are not optimized.

This libray is what Toy - MNIST Classification and Toy - Generate text with character-based RNN are based on.

Directories

• neural/ - contains the basic building blocks that can be used to build and train artificial neural networks.
• examples/mnist/ - MNIST classification example.
• examples/text_generation/ - Text generation with character-based RNN
• ndarray/ - multi-dimensional arrays a la numpy
• trees/ - functions to manipulate tree-like structures a la Pytrees
• arrays/ - array utilities
• floats/ - Float32Array utilities
• numerical/ - mathematical utilities