www.Autodiff.org - Publication: Demystifying differentiable programming: shift/reset the penultimate backpropagator

Demystifying differentiable programming: shift/reset the penultimate backpropagator

- Article in a journal -

Author(s)
Fei Wang , Daniel Zheng , James Decker , Xilun Wu , Grégory M. Essertel , Tiark Rompf

Published in
Proceedings of the ACM on Programming Languages

Year
2019

Publisher
Association for Computing Machinery

Abstract
Deep learning has seen tremendous success over the past decade in computer vision, machine translation, and gameplay. This success rests crucially on gradient-descent optimization and the ability to “learn” parameters of a neural network by backpropagating observed errors. However, neural network architectures are growing increasingly sophisticated and diverse, which motivates an emerging quest for even more general forms of differentiable programming, where arbitrary parameterized computations can be trained by gradient descent. In this paper, we take a fresh look at automatic differentiation (ad) techniques, and especially aim to demystify the reverse-mode form of ad that generalizes backpropagation in neural networks. We uncover a tight connection between reverse-mode ad and delimited continuations, which permits implementing reverse-mode ad purely via operator overloading and without managing any auxiliary data structures. We further show how this formulation of ad can be fruitfully combined with multi-stage programming (staging), leading to an efficient implementation that combines the performance benefits of deep learning frameworks based on explicit reified computation graphs (e.g., TensorFlow) with the expressiveness of pure library approaches (e.g., PyTorch).

AD Theory and Techniques
Functional Programming

BibTeX
@ARTICLE{
         Wang2019Ddp,
       author = "Wang, Fei and Zheng, Daniel and Decker, James and Wu, Xilun and Essertel,
         Gr\'{e}gory M. and Rompf, Tiark",
       title = "Demystifying differentiable programming: shift/reset the penultimate backpropagator",
       year = "2019",
       issue_date = "August 2019",
       publisher = "Association for Computing Machinery",
       address = "New York, NY, USA",
       volume = "3",
       number = "ICFP",
       url = "https://doi.org/10.1145/3341700",
       doi = "10.1145/3341700",
       abstract = "Deep learning has seen tremendous success over the past decade in computer vision,
         machine translation, and gameplay. This success rests crucially on gradient-descent optimization and
         the ability to “learn” parameters of a neural network by backpropagating observed
         errors. However, neural network architectures are growing increasingly sophisticated and diverse,
         which motivates an emerging quest for even more general forms of differentiable programming, where
         arbitrary parameterized computations can be trained by gradient descent. In this paper, we take a
         fresh look at automatic differentiation (AD) techniques, and especially aim to demystify the
         reverse-mode form of AD that generalizes backpropagation in neural networks. We uncover a tight
         connection between reverse-mode AD and delimited continuations, which permits implementing
         reverse-mode AD purely via operator overloading and without managing any auxiliary data structures.
         We further show how this formulation of AD can be fruitfully combined with multi-stage programming
         (staging), leading to an efficient implementation that combines the performance benefits of deep
         learning frameworks based on explicit reified computation graphs (e.g., TensorFlow) with the
         expressiveness of pure library approaches (e.g., PyTorch).",
       journal = "Proceedings of the ACM on Programming Languages",
       month = "jul",
       articleno = "96",
       numpages = "31",
       pages = "96:1--96:31",
       keywords = "Automated Differentiation, Delimited Continuations, Differentiable Programming,
         Multi-stage Programming",
       ad_theotech = "Functional Programming"
}

back

autodiff.org
Username:
Password: