ABSTRACT
We develop an approach to measure the statistical performance in learning from
noisy examples, so as to draw a learning curve. Learning is quantified with a
distance-like function, varying with the number of examples presented. Upper
and lower benchmarks for the real learning curve can be considered,
corresponding to an ideal observer and, to the other extreme, to a mere random
choice. For the ideal observer, we consider some possible cognitive strate-gies
for the identification of the stimuli, differing in the role of the features
comprising the complex stimuli and in the amount of prior knowledge
exploited. This theoretical framework can provide a tool for analysing
experimental human performance in appropriate psychophysical tests.
