Seminar on History and Philosophy of Science

Abstract: Imagine that you are a scientific project leader. Following what many scientists have done, you are thinking about "hiring" an (idealized) Bayesian agent---that is, using a computer to simulate one Bayesian agent or another to help you infer from data to one of the competing hypotheses on the table. Setting computational cost aside, which Bayesian agent(s) are the best candidates for you and your project? This is basically the so-called problem of the priors in Bayesian epistemology (although the term 'prior' has not made its appearance). The most familiar solution is called subjective Bayesianism, whose slogan is: "Coherence alone is good enough." But I want to articulate and defend an alternative solution, whose slogan may be formulated as: "Coherence is great but a carefully planned pursuit of truth is no less important." This alternative solution has gradually become influential among Bayesian statisticians, and some of its core ideas have long been standard in the broader scientific community (keyword: 'statistical consistency'). But, ironically, it has remained quite underappreciated, both in science and in philosophy. One main reason, I think, is that it is often confused with a special version of subjective Bayesianism (which appeals to a Bayesian norm called regularity). I defend the "pursuit of truth" solution by doing two things: (i) giving it a clear formulation that it has long deserved, and (ii) replying to various worries. This talk will not, and need not, presuppose background knowledge in statistics---for the defended view actually applies not just to statistical inference but to scientific inference in general.