The gap between abstract and concrete results in machine learning
Peter D. Turney · Journal of Experimental & Theoretical Artificial Intelligence · 1991
The gap between abstract and concrete results in machine learning is largely due to unrealistic assumptions made by researchers in formal learning theory. The task of abstractly describing learning is very difficult, so it is natural to make simplifying assumptions. As research progresses, assumptions will slowly become more realistic. Current work in formal learning theory has not yet advanced to the stage where it will be useful to empirical and psychological research in machine learning. Future research in formal learning theory must attack the problem of bias. Bias is essentially a preference ordering on hypotheses. Two forms of bias are suggested here. Recent research in machine learning has examined algorithms that shift their bias for different domains. The biases discussed here resist this trend, as they attempt to apply to all ‘natural’ domains.