deleted

I agree with the original poster. Offering MXNet alternatives needs to be a very high priority. Perhaps doing so would help with the significant problem of Target->"GPU" that at least of version 13.3 did not work well on Macs when training neural nets.

Btw, regarding Jax, I think of Mathematica implemented tensor differentiation natively, this would obviate the need to rely on external autodiff.

The issue is that neural network training involves differentiation of expression like norm(W1 W2 W3) w.r.t to matrix W2. Mathematica differentiation is at its core a scalar differentiation, it can't treat W2 as an atom. Autodiff frameworks like Jax can find these derivatives efficiently by treating matmuls as atomic operations.

I have thought a lot about the proper design of such a system and have a couple of prototypes, getting it into Mathematica is on my to-do list https://community.wolfram.com/groups/-/m/t/2437093

Just wanted to alert that as of mid 2023 Google is recommending Flax for new projects instead of Haiku

Jax (and PyTorch) are already "integrated" with Mathematica using Python ExternalFunction interface, so the question would be -- which parts of Jax/PyTorch would you want to integrate more closely?

If the goal is to replace MxNet parts impementing NN interface, then it would be the modeling code.

For PyTorch, that would mean integrating nn.Module abstraction which gives is a building block that you wire up together with other nn.Module blocks to build your neural net. This abstraction has been around for 8+ years, so it probably won't go away soon. For instance, models from PyTorch model hub and most repos under Papers with Code use nn.Module interface. So one interesting use-case would be to make give a way to load model hub models into a Mathematica symbol which could be called as a function.

Jax has started as "autodiff for numpy", so neural network abstractions didn't come until later. There's Flax created by Google and Haiku that was created by DeepMind. Integrating Jax modeling would mean deciding whether you want to integrate Flax Dense or Haiku Linear layers.

We are well aware of the MXNet situation and have been planning a backend switch. It is a very large project that's currently in its early research state, and at the moment we don't yet have a time estimate for its completion.

What machine learning framework (or frameworks) are of interest to you (as a possible complete replacement for MXNet)?

Building on Yaroslav Bulatov's post:

Closely related is support for things like NeuralODEs (backpropogating through NDSolve). This is another increasingly important area in scientific ML, and is now built into Julia (https://julialang.org/blog/2019/01/fluxdiffeq/) and PyTorch (https://towardsdatascience.com/neural-odes-with-pytorch-lightning-and-torchdyn-87ca4a7c6ffd) . One can kind of "hack it" by implementing ODE integrators as neural nets in Mathematica (I've done it, as have some others).