HANDLING SMALL TRAINING SETS CONFIDENCE/ACCURACY WITH REGARD TO NEW EXAMPLES

摘要

It often happens that the number of samples available to train a discriminator is many fewer than Learning Theory tells us we need to accomplish the required accuracy/confidence. When you run up against a theoretical limit, only two choices are possible. You can accept the situation, or you can look for ways around those limits. This report suggests that there is a way around conventional learning theory and applies the new technique (called Margin Setting") to a difficult artificial problem to illustrate its power.