Neural Network Model File Format

I’ve uploaded a README [1] to cover the ongoing work on devising a simple file format for neural network models along with a small JavaScript library for use with WebNN [2], W3C’s platform neutral API for running neural networks on NPUs, GPUs or CPUs.  Whilst there are existing libraries and model formats, they are HUGE and hard to work with.  A further motivation is that WebNN is intended for inference, not training.  The new library will transform the inference model into an inverse training model suitable for use with WebNN.

This work is still at an early stage, with the current focus on ensuring the file format is adequate for typical neural network models. The plan is to first to support running the inference model, and then to work on generating the training model, starting with simple models like a two layer perceptron, before extending support to recurrent models which need to be unrolled to work on WebNN.

My goal is to work towards applying web-based federated learning to enable virtual presence in the immersive Web. If you haven’t read it before, you may enjoy how this is foretold in Neal Stephenson’s 1992 cyberpunk novel “Snow Crash”.

[1] https://github.com/w3c/cogai/blob/master/nnm/README.md
[2] https://www.w3.org/TR/webnn/
 
Best regards,

Dave Raggett <dsr@w3.org>