Sorry, I misspoke - will edit. Was talking about resistance to overfitting. Which largely comes from logistic regression's assumption of a linear decision bound

Sorry, I misspoke - will edit. Was talking about resistance to overfitting. Which largely comes from logistic regression's assumption of a linear decision boundary. It's true surprisingly often in classification tasks, and, when it's not, you can usually model it just fine with interaction variables. With an ANN, your easiest defense against overfitting is to have great big heaping piles of training data. That's something that's hard to come by in many interesting situations.