matnets 3.0.0

A small experimental neural network library where neurons are represented as matrices.

MATNETS

MATNETS is a JAX library for matrix-neuron neural network experiments.

In a traditional neural network, a neuron carries one scalar activation. In MATNETS, each neuron carries an n x n matrix. A layer maps a stack of input matrix-neurons to a stack of output matrix-neurons.

Core Shape Contract

The core dense primitive uses:

params.W: (q, p, n, n)
params.B: (q, n, n)
x:        (p, n, n)
output:   (q, n, n)

p is the input neuron count. q is the output neuron count. n is the matrix size for every neuron.

Read Next

• Getting Started: install MATNETS and run a first dense layer.
• Concepts: understand matrix-neuron shapes and JAX transforms.
• API Guide: see each public function and its expected shapes.
• Examples: run the included examples.
• Development: run tests and local checks.