Timezone: »
We review the current state of automatic differentiation (AD) for array programming in machine learning (ML), including the different approaches such as operator overloading (OO) and source transformation (ST) used for AD, graph-based intermediate representations for programs, and source languages. Based on these insights, we introduce a new graph-based intermediate representation (IR) which specifically aims to efficiently support fully-general AD for array programming. Unlike existing dataflow programming representations in ML frameworks, our IR naturally supports function calls, higher-order functions and recursion, making ML models easier to implement. The ability to represent closures allows us to perform AD using ST without a tape, making the resulting derivative (adjoint) program amenable to ahead-of-time optimization using tools from functional language compilers, and enabling higher-order derivatives. Lastly, we introduce a proof of concept compiler toolchain called Myia which uses a subset of Python as a front end.
Author Information
Bart van Merriënboer (MILA, Google)
Olivier Breuleux (MILA)
Arnaud Bergeron (Université de Montréal (Mila))
Pascal Lamblin (Google)
Related Events (a corresponding poster, oral, or spotlight)
-
2018 Spotlight: Automatic differentiation in ML: Where we are and where we should be going »
Thu. Dec 6th 03:40 -- 03:45 PM Room Room 220 CD
More from the Same Authors
-
2019 Workshop: Program Transformations for ML »
Pascal Lamblin · Atilim Gunes Baydin · Alexander Wiltschko · Bart van Merriënboer · Emily Fertig · Barak Pearlmutter · David Duvenaud · Laurent Hascoet -
2018 Poster: Tangent: Automatic differentiation using source-code transformation for dynamically typed array programming »
Bart van Merriënboer · Dan Moldovan · Alexander Wiltschko -
2017 Workshop: The future of gradient-based machine learning software & techniques »
Alex Wiltschko · Bart van Merriënboer · Pascal Lamblin -
2016 Workshop: The Future of Gradient-Based Machine Learning Software »
Alex Wiltschko · Zachary DeVito · Frederic Bastien · Pascal Lamblin -
2014 Demonstration: Neural Machine Translation »
Bart van Merriënboer · Kyunghyun Cho · Dzmitry Bahdanau · Yoshua Bengio -
2007 Poster: Learning the 2-D Topology of Images »
Nicolas Le Roux · Yoshua Bengio · Pascal Lamblin · Marc Joliveau · Balázs Kégl -
2006 Poster: Greedy Layer-Wise Training of Deep Networks »
Yoshua Bengio · Pascal Lamblin · Dan Popovici · Hugo Larochelle -
2006 Talk: Greedy Layer-Wise Training of Deep Networks »
Yoshua Bengio · Pascal Lamblin · Dan Popovici · Hugo Larochelle