Reliable Reasoning, by Gilbert Harman and Sanjeev Kulkarni.
John Williamson · Mind · 2012
The aim of the book is to give a non-technical introduction to statistical learning theory at undergraduate level. Statistical learning theory is concerned with the reliability of rules for classifying a new case — e.g. diagnosing a disease in a new patient — on the basis of other features of the case and a large stock of past cases and their features and classifications. The book is based on a course on learning theory and epistemology given to undergraduate students in electrical engineering and in philosophy at Princeton. It is a short book, with four chapters. The first chapter is on ‘the problem of induction’. Note though, that the book is not concerned with what philosophers normally take to be ‘the problem of induction’ — the problem of justifying induction — but rather with the problem of assessing the reliability of inductive rules. The chapter then sets out to debunk the commonly held view that there are two sorts of reasoning, deductive and inductive, and two sorts of arguments, deductive and inductive, by means of arguments to which I will return below. Chapter two introduces the idea of using enumerative induction to learn classification rules, and for estimating the values of continuous variables. It introduces the VC-dimension of a set of classification rules, which is perhaps the most important concept in statistical learning theory. Chapter three discusses induction rules which work by trying to rank hypotheses by their simplicity. Chapter four discusses applications of statistical learning theory to neural networks and support vector machines in machine learning.