Perceptron overfitting? I'm trying to judge the performance of my perceptron linear discriminant. In one instance I'm training on a sample size of 150 and on another I'm training on a sample size of 1500. I test both of these on the same testing set. Yet, the peceptron that is trained on the larger set performs more poorly. I'm not sure how to describe this...could this be an example of overfitting??
 A: The original perceptron algorithm goes for a maximum fit to the training data and is therefore susceptible to over-fitting even when it fully converges. You are also right in being surprised, because when the number of training data increases, over-fitting usually decreases. However, the original perceptron can over-fit even to one single noisy example in your training data, and sometimes such adverse examples are rare and tend to appear only in a larger set. That can be one justification for you observation. 
Another justification can be that your algorithm did not converge on one or both of the data sets, and the resulting weight vectors are of low quality (you did not specify convergence). This is a major weakness of the original perceptron algorithm. In any case, trying the algorithm on more data sets (to see if there is a pattern) and/or using the averaged or large-margin perceptrons can help. 
A: Assuming that there is no sampling bias in your training sets, I would test it using different sample sizes, e.g. 100, 200, ..., 1500 to see how the performance evolves. If at some point your performance starts to decrease then it is a pretty good indicator it's starting to overfit.
