Extending the Valiant framework to detect incorrect bias

Since it is difficult to know the correct bias for an inductive learning problem a priori, the ability to detect a bad bias can be valuable. One method is to take advantage of an algorithm that makes strong performance guarantees when the bias is correct, then verify that the algorithm performs as promised. This paper develops this idea within the context of the Valiant framework. In the basic Valiant framework, / / the assumption that the target concept belongs to a given concept class holds then the output of a learning algorithm is (1 <$)-reliable. After incorporating the notion of bias-evaluation, the assumption is removed such that the output is (1 <5)-reliable regardless of whether or not the target concept belongs to the given concept class. It is shown how to convert an existing pac-learning algorithm into one with reliable bias-evaluation. This research was sponsored by NASA under contract number NAGW-1175. The views and conclusions contained in this document are those of the author and should not be interpreted as representing the official policies, either expressed or implied, of NASA. Extending the Valiant Framework to Detect Incorrect Bias Lonnie Chrisman