Learning Speed Curves: Prediction of Average Case Learning Using VC-Dimension Analysis and Regression
The objective of this project is to predict the learning speed curve for an inductive learning algorithm, when given just a small number of examples drawn from the target distribution. This objective differs form most existing research in that the goal is to predict average case performance, not worse case performance; and to produce results for both noisy and noise-free input.
Existing general regression techniques are analyzed with respect to their
ability to accurately create
predictive learning-speed curves.
A new method of general regression is presented, and implemented in a system
called SEER. The new method’s
model, called the Effective Dimension Model, is based on the Vapnik-Chervonenkis
dimension.
The described experimental results show that SEER accurately predicts, from a
small sample of cases, with and without noise, the number of cases required to
achieve a desired level of classification accuracy.