[ programme | | outline | short biography ]
JEAN-NICOD LECTURES 2005
Gilbert HARMAN
The Problem of Induction and Statistical Learning Theory
Outline of the lecturesLecture 1: The Problem of Induction.
In assessing the reliability of inductive inferences, it is important not to think of induction and deduction as two kinds of inference, because a mistake to think that there is such a thing as deductive inference. The fact that inductive reasoning often leads one to give up things previously believed may seem to make it hard to specify what reliability comes to, but in fact developments in statistical learning theory allow a way to specify a kind of enumerative induction and answer certain questions about its reliability and about the reliability of certain other inductive methods.
[Full Text]Lecture 2: Enumerative Induction in Statistical Learning Theory and Popper on Falsifiability.
A certain sort of enumerative induction which selects from a limited set of rules that rule that best fits the data. In the theory of machine learning, there is a precise statement of the conditions under which reliable enumerative induction of this sort is possible, namely, it is possible only if the set of hypotheses being considered is not too rich, where richness is inversely correlated with falsifiability in something like Popper's sense. In this lecture I describe some of the relevant theory and begin to discuss Popper's view.
Lecture 3: Going Beyond Enumerative Induction.
A different kind of induction balances data-coverage against something like simplicity. One criterion might be the the length of a statement of the hypothesis. Another idea would be to allow infinite classes of equally simple hypotheses, so that all linear hypotheses go into one class, for example, and the complexity of a class of hypotheses is measured by the number of parameters needed to specify a particular instance of the class. Such a measure gives bad results for trigonometric functions, however, and it is necessary instead to appeal again to falsifiability, this time to measure complexity.
Lecture 4: Support Vectors and Transduction.
Recent developments include use of support vector machines and methods of induction that infer directly from data to the classification of new cases as they have come up, without basing the classification on the prior acceptance of a general rule. Developments in statistical learning theory raise questions about realism versus instrumentalism, about basic science versus applied science, that is, about when to try to find an underlying rule and when to forget that and try to reach conclusions directly about the next instance.
Further information:
Sophie Bilardello
Institut Jean-Nicod
1bis avenue de Lowendal
75007 Paris
phone : 01 53 59 32 81
fax : 01 53 59 32 99