Neural Network Learning: Testing Bounds on Sample Complexity

Several authors have theoretically determined distribution-free bounds on sample complexity. Formulas based on several learning paradigms have been presented. However, little is known on how these formulas perform and compare with each other in practice. To our knowledge, controlled experimental results using these formulas, and comparing of their behavior, have not so far been presented. The present paper represents a contribution to filling up this gap, providing experimentally controlled results on how simple perceptrons trained by gradient descent or by the support vector approach comply with these bounds in practice.