First draft for WebNNM spec

Existing neural network frameworks are huge and hard to use. WebNNM is a lightweight easy to understand high-level framework that simplifies working with neural networks for newcomers whilst offering experts fine grained control.

I’ve uploaded an initial draft for the Web Neural Network Models specification:

See: https://w3c.github.io/cogai/WebNNM/

I expect to refine it, but even this draft gives an indication of the breadth of the library.   I am currently working on stabilising training when using float16, something desirable for increased speed and reduced power consumption on consumer grade devices, e.g. Apple devices with the ANE.  I then need to finish the test harness and expand the coverage of the WebNN operators.

In the meantime take a look at the demos: https://github.com/w3c/cogai/blob/master/WebNNM/README.md

p.s. I gratefully acknowledge the use of Google’s Gemini LLM which has proved invaluable when working on the design and implementation.  LLMs are great when you figure out what to ask, but not so good at critical independent thought.

Dave Raggett <dsr@w3.org>