14 ms·

Regarding the perceptron, most modern texts also have a regret analysis for the perceptron, which coupled with an online-to-batch conversion tells you how well

by alextp 13y ago

Regarding the perceptron, most modern texts also have a regret analysis for the perceptron, which coupled with an online-to-batch conversion tells you how well do you expect a perceptron to perform on unseen data after a single pass, and it's usually a very good estimate (the answer is on average about as well as it did on the examples in the training data).

syllogism 13y ago
This isn't often discussed in my field (NLP) -- thanks! We usually use averaged perceptron though, which seems like it would make this hard to analyse.
- alextp 13y ago
  The averaged perceptron is the one to which the proof applies :-)