There have been some past discussions about this:
You can implement various forms of automatic differentiation in the language:
-
Using dual numbers: https://mathematica.stackexchange.com/questions/1873/is-there-any-automatic-differentiation-package
-
Using tensor networks: Tensor networks, Einsum optimizers and the path towards autodiff
If you are referring to autodifferentiation of neural networks…the current system (as of 14.0) based on MXNet does not seem to support doing this.
- Discussion on StackExchange: https://mathematica.stackexchange.com/questions/256323/neural-network-automatic-differentiation-autograd
- Discussion about MXNet and its limits: About MxNet as Mathematica backend choice