Principles of Learning Problem Design

Things have been a little quite around here of late, mainly because I’ve been working on a submission for the NIPS 2007 Workshop on Principles of Learning Problem Design in early December.

I’m pleased to say that I’ll be presenting some recent results that Bob and I have been working on under the heading of “Representations in Learning Task Design”. The focus is on finding primitives and combinators for describing learning tasks (Aside: “problems” are what you are trying to solve, “tasks” are what you give to computers to solve them).

Unsurprisingly, cost-sensitive losses are one such primitive and when combined using weighted integration they can represent a variety of losses for a range of learning tasks including classification, regression and class probability estimation.

Since this is a workshop paper, most of the results are still fairly preliminary and build on a lot of work by others. That said, I think it’s a good approach as several previously known results are subsumed and simplified. I’ll post our paper and slides once they are completed.

Let me know if you are attending NIPS and we can try to catch up. Hope to see you at Whistler.

Mark Reid November 20, 2007 Canberra, Australia
Subscribe: Atom Feed