I believe the right question to ask is whether we need domain experts at all, or just need a whole lot of data.

I think the answer’s pretty obvious. Even the basic structure of a statistical model entails a large degree of design in everything from setting up dependencies to selecting predictors.

The most accurate natural language systems bring in all kinds of human-generated knowledge sources from labeled data for classifiers or part-of-speech taggers to domain-specific dictionaries to full-blown ontologies.

I’d cut just about anybody slack for not sorting out all of our redundant terminology. I only just recently realized that so-called max entropy classifiers, logistic regression, and one-layer neural nets with sigmoid/softmax activation were the same thing, and that L1 norms, Laplace priors, double-exponential priors, and the “lasso” are the same thing.