Seminar on Computational Learning and Adaptation

  Let's Improve Machine Learning

Oliver Selfridge
MIT Media Lab
BBN Technologies
Boston, Massachusetts

Abstract:

Today, in Artificial Intelligence (AI), Machine Learning is a vigorous
and flourishing field. I believe that we can and ought to do more. My
overall recommendation for ML is that now we should find out how to
produce cognitive software that can be at least partly educated
instead of having to be carefully programmed. The software must be
able to learn not only how to accomplish the top level desired task,
but also how to check and improve its performance on a continuing
basis at many different levels.

There are four main topics in human learning that are mainly not even
considered in most of ML. The first is what I have termed purpose
structure; which means that software should care! The idea of purpose
structures is to build software out of modules each of which has a
success function, so that changes in them can be assessed to assure
continuing improvement. The second topic is: how are the conclusions
of ML in a piece of cognitive software to be remembered, so that what
has been learnt can be applicable again in later and perhaps different
circumstances? The third topic is that anything learnt by people is
rarely handled as an isolated and independent piece of knowledge;
rather, it is embedded in a structure of some conceptual models. The
fourth topic is: how are the conclusions of ML in a piece of cognitive
software to be shared with other cognitive agents?...what kind of
languages should be used? for most of what we know we learnt from
others, not from our own experiences.

None of those general topics has been much faced in AI, let alone in
ML. On top of that, the cognitive software must work in environments
that are continually changing at all levels, including the overall
standards of success. We need to analyze those points and put them in
some kind of order so as to be able to analyze and attack them. Then
we can propose a program that will diverge and then we can take the
one less traveled by perhaps that will make all the difference! And
perhaps we can then break new boundaries in AI.

This talk is cosponsored with the Symbolic Systems Forum.

Return to the seminar schedule