[PRL] Fwd: Whatever happened to partial evaluation?

Matthias Felleisen matthias at ccs.neu.edu
Mon Nov 28 14:37:53 EST 2011


On Nov 28, 2011, at 1:56 PM, Aaron Turon wrote:

> The reason I ask is that my group is trying to get automatic
> differentiation to be both first-class and efficient.  (Why?  Because
> AD can generate gradients of machine learning models without manual
> input -- if made practical, this could revolutionize machine
> learning.)  Efficient but not first-class, and in particular poorly
> nestable, automatic differentiation can be had from the
> source-to-source transformers at autodiff.org; first-class but
> inefficient automatic differentiation can be had by operator
> overloading. 


Hi Alexey, I cannot respond to your general question but I
would like to point to two people re your specific research
agenda. 

-- Jeff Siskind (Purdue) and collaborators worked on just this idea. 
See their POPL '07 papers in Nice, France. They work in Stalin Scheme, 
and they are definitely aware of partial evaluation. 

-- Mike Fagan (Rice) is the main author of a 90-ish award winning 
differentiation program (they also dealt with adjoins) for Fortran, 
ADIFOR. The first version of this program was also written in Scheme, 
and for all I know they still use Scheme. Mike is definitely familiar
with partial evaluation and such things. 

You may wish to ask these two people because they can evaluate your 
question in context. 

-- Matthias




More information about the PRL mailing list