Re: Differentiable Forth

On Fri, 19 Jul 2024 19:19:24 +0000, ahmed wrote:

On Fri, 19 Jul 2024 18:01:39 +0000, minforth wrote:
>

I don't know if gforth has z: locals for complex numbers. If yes, they
could be used for dual fp-numbers as well to reduce code overhead and
improve readability.

>
Yes, z: locals exists in gforth but I haven't used it. I didn't want to
mix dual and complex numbers.
I have completed the code with the elementary functions and their
derivatives (exp, ln, trigs, inverse trig, hyperbolic, inverse
hyperbolic, ...).
The code is not optimized and can be improved. I've done it just as a
proof of concept.

Thanks for sharing this concept. It is often surprising in how few
lines of Forth code one can express seemingly complex topics.
Automatic differentiation, especially in its reverse form, is a
fundamental algorithm for e.g. machine learning, thereby processing huge
sparse matrices. Forth is lightyears away from the domains of Jax and
TensorFlow. Nevertheless, I believe that if Forth makes any progress
at all, it should be in the direction of better data handling. You can
do only so much with one or two stacks ...