Barak A. Pearlmutter | 15 Jan 13:01 2014

Postdocs / Research Programmer for Efficient First-Class Automatic Differentiation

                     Postdocs / Research Programmer
   Compositional Learning via Generalized Automatic Differentiation

The goal of this project is to make a qualitative improvement in our
ability to write sophisticated numeric code, by giving numeric
programmers access to _fast_, _robust_, _general_, _accurate_
differentiation operators.

To be technical: we are adding exact first-class derivative
calculation operators (Automatic Differentiation or AD) to the lambda
calculus, and embodying the combination in a production-quality fast
system suitable for numeric computing in general, and compositional
machine learning methods in particular.  Our research prototype
compilers generate object code competitive with the fastest current
systems, which are based on FORTRAN.  And the combined expressive
power of first-class AD operators and function programming allows very
succinct code for many machine learning algorithms, as well as for
some algorithms in computer vision and signal processing.  Specific
sub-projects include: compiler and numeric programming environment
construction; writing, simplifying, and generalising, machine learning
and other numeric algorithms; and associated Type Theory/Lambda
Calculus/PLT/Real Computation issues.

TO THE PROGRAMMING LANGUAGE COMMUNITY, we seek to contribute a way to
make numeric software faster, more robust, and easier to write.

TO THE MACHINE LEARNING COMMUNITY, in addition to making it easier to
write efficient numeric codes, we seek to contribute a system that
embodies "compositionality", in that gradient optimisation can be
(Continue reading)