Foundations for Efficient and Expressive Differentiable Programming [pdf]

http://papers.nips.cc/paper/8221-backpropagation-with-callbacks-foundations-for-efficient-and-expressive-differentiable-programming.pdf

by espeed • 8 years ago

80 14 8 years ago

papers.nips.cc

tiarkrompf5• 8 years ago

Co-author here. Happy to answer any questions, as usual ...

tiarkrompf5• 8 years ago

We also have a more recent and slightly longer draft with additional explanations and GPU training results for ResNet 50 and DeepSpeech2:

"Demystifying Differentiable Programming: Shift/Reset the Penultimate Backpropagator" https://www.cs.purdue.edu/homes/rompf/papers/wang-preprint20...

The Lantern framework is available here: https://github.com/feiwang3311/Lantern

improbable222• 8 years ago

This looks interesting, thanks.

What you call "delimited continuations" sounds a bit like how AD works in Julia's Flux package (and maybe elsewhere): During the forward calculation, a chain of functions is constructed, whose evaluation is the backward pass. This is done by overloading the original * to return both x * y and a closure Δ -> (Δ * y, x * Δ).

Does that sound right, are these indeed similar, or have I mis-understood something? I have never read Scala, but if I squint I can make your overloading of * is doing something similar.

tiarkrompf5• 8 years ago

It's different -- instead of returning a closure, we take a closure as additional parameter (check paper for details). This means that the call stack stays intact and all intermediate values can be stack-allocated.

cs702• 8 years ago

Very cool.

Like all really good ideas, this one seems "obvious" in hindsight. I mean that is a compliment: It would have never occurred to me that transforming code into continuation-passing-style code would allow for automatic differentiation through all dynamic control-flow structures, by leveraging the function-call stack, thus eliminating the need for some kind of "tape" data structure, e.g., as in PyTorch.

My question is about the ongoing work to provide a JIT compiler for Python code. Do you expect it will provide full support for the entire PyTorch and/or Tensorflow APIs?

tiarkrompf5• 8 years ago

Thanks! Yes, it wasn't obvious at all when we started looking at AD either.

Lantern supports a good deal of PyTorch (via Snek, our Python front-end similar to AutoGraph) and can also read ONNX. Full feature parity is not our main goal--so far, supported features have been driven mostly by what is required for certain interesting models.

carterschonwald• 8 years ago

Another paper on the same topic that got best paper at icfp this fall is http://conal.net/papers/essence-of-ad/

Which shows how to connect forward and reverse mode auto differentiation in a very elegant way

mlazos• 8 years ago

I saw this paper at nips and I thought it was an awesome application of continuation passing style. Typically in CPS one of the advantages is that you really only need a single stack frame (since the remainder of the program is encoded in the callback), but in this paper they allow the CPS’d program to use multiple stack frames and store the intermediate results from backprop in the callstack. It never would’ve occurred to me to remove one of the advantages of CPS in order to store values on the callstack. Cool idea!

Foundations for Efficient and Expressive Differentiable Programming [pdf]

Related Stories