Richard Wei

> Could I do something like this (this is a hastily written sample only): https://gist.github.com/tanmayb123/c57c956120d3733ed0691c6406a74284 This style can work, but it is not recommended. We recommend using recursive algebraic data...

Backpropagation doesn't necessarily require control flow differentiation once you are comfortable using pullbacks. Here's an RNN example: https://gist.github.com/rxwei/ce6644efad8f229651050e096c05ccbb.

Thank you for filing the issue! Before we go about implementing it in the standard library, it is not 100% clear this is the right direction, and thus some prototyping...

Hi @tanmayb123, let’s not add those operator-like layers for now and defer it to further discussions. Ideally, we would want functions to be able to conform to the layer protocol,...

Random thought: We can define layer wrappers for common arities (unary and binary), and define a `layer(_:)` factory function for turning any `Differentiable`-to-`Differentiable` function into a layer. Rough sketch: ```swift...

@Shashi456 Thanks a lot for listing these! Looks good to me. I'd suggest starting with the non-recurrent ones first.

Absolutely! What would you like to implement?

Sounds great. Look forward to your PR.

IMO it is ideal to stick with one set of names for consistency in all our models and example code. Currently we are leaning towards consistency with Keras. This will...

We already have such layers, `Reshape`, for example. Adding a layer wrapper for each function is definitely not ideal and would complicate our API surface. Instead of throwing a lot...