Relations between machine learning problems









Relations between machine learning problems – an approach to unify the field

NIPS 2011 Workshop, December 16 or 17, 2011, Sierra Nevada,Spain

The only essential knowledge pertains to the
inter-relatedness of things
--- Jorge Luis Borges

This workshop will focus on relations between machine learning problems. The idea is that by better understanding how different machine learning problems relate to each other, we will be able to better understand the field as a whole.

The idea of a relation is quite general. In includes such notions as reductions between learning problems, but is not restricted to that. Our goal can be explained by an analogy with functional analysis - rather than studying individual functions, functional analysis focusses on the transformations between different functions. This high level of abstraction led to enormous advances in mathematics.

The motivation for the workshop is several-fold:
  • End users typically only care about solving their problem, not the technique used. Many machine learning techniques still require a detaile dunderstanding of their operation in order to use them
  • ML as a service Much modern software is evolving to being delivered via the web as a service. What does it mean  for Machine Learning to be delivered as a service? One question that needs resolving is how to describe what the service does (ideally in a declarative manner). Understanding relations between machine learning problems can be thus seen as analogous to the composition of (Machine Learning) web services
  • Reinvention Many machine learning solutions are reinvented / rediscovered. This is hardly surprising since the focus is often on techniques and not problems. If you can not describe your problem in a manner that others can easily understand and search, then it is hard to figure whether solutions to seemingly new problems already exist.
  • Modularity A feature of mature engineering disciplines is modularity, which has enormous design and economic advantages. Understanding relations between problems seems important to achieve greater modularity.
  • Conceptual simplicity Finally, if one can understand the field using a smaller number of primitives and combination operations, then this has an intrinsic appeal (apply Occam's razor at the meta-level!)

Here is a more detailed description.

We aim to prepare a position paper by mid-September that better summarises our thinking and the possible directions forward.


  • Submissions     30 September, 2011
  • Notifications      23 October, 2011
  • Registration       (See nips website)


Bob Williamson
John Langford
Ulrike von Luxburg
Mark Reid
Jennifer Wortman Vaughan


  • Introductory overview linked to the position paper (the organisers)
  • We need a BIT more GUTS (=Grand Unified Theory of Statistics)  (Peter Grünwald)
  • Machine Learning Markets: Putting your money where your mouth is  (Amos Storkey)
  • Contributed talks
  • Panel discussion: How to build a map of all of machine learning? 

Call for submissions


We will seek funding to enable the talks to be recorded and published by

We do not intend to publish workshop papers  (we leave it to individual authors to choose the best venue for their final papers).


For enquiries and submissions email