VLADIMIR VAPNIK (Royal Holloway, London)

Problem of empirical inference in machine learning and Philosophy of Science

1.   The problem of Empirical Inference is a core problem of human intelligence.  It has been discussed for more than 2,000 years.  However only since the 1960s with the appearance of fast computers did it become a full-edged subject of Natural Science (as Physics or Biology).  Now one can not only speculate about models of learning, but can also conduct wide scale experiments with computers.

Such experiments have demonstrated that many prejudices were accumulated during the time when Empirical Inference Science was driven mostly by speculation.  Since the 1970s Machine Learning has made great progress.  In particular, complete answers to core mathematical problems of generalisation were found.

2.   In my talk I will discuss results of the Mathematical Learning Theory from general point of view of Philosophy of Science.  I will restrict myself mostly to facts that have the status of necessary and sufficient conditions and, therefore, must be satisfied by any learning system (including humans).  I will try to show the restrictiveness of understanding learning problems in humanist studies and in particular in classical Philosophy of Science.

3.   A crucial point in Machine Learning Science was the discovery (both in theory and in experimental studies) of the existence of direct non-inductive methods of inference (from data to data, avoiding a generalization stage).  It has been proven that non-inductive inferences are always more accurate than inductive.  It has been shown that there are exist a large family of non-inductive methods.

4.   The problem of Empirical inference - along with pure mathematical concepts - contains concepts that have clear humanist interpretations (for example "cultural Universum") that are important instruments for effective inferences.  I will try to show that the problem of Empirical Inference has reached the point where progress requires the joint efforts of philosophers and statisticians. In particular, my main thesis will be that any breakthrough in understanding learning will require primarily philosophical (conceptual) rather than mathematical (technical) advances.

Vladimir's pdf slides