Colloquium: "A Molyneux Problem for Automated Science"

Abstract: Automated science is the attempt to use a combination of machine learning and robotics to perform experiments, build new models, postulate new laws, and design new experiments on the basis of what is observed to be the result of previous experiments. The comparison between the learning problems faced by automated machine learning and the learning problems faced by human scientists sheds light on a subtle measurement problem which I will call the Molyneux Problem for Science.  The Molyneux problem asks whether a newly sighted person could recognize objects previously known by touch. I generalize the Molyneux problem to all forms of data and characterize the problem as a special form of scientific learning. The Molyneux Problem for Science exists whenever identity must be recognized across perceptual modalities or data streams of different types without the aid of causal or correlational information.  After distinguishing the scientific Molyneux problem from inverse problems and problems of causal discovery, I will use the case of early radio telescopy to demonstrate that scientific Molyneux problems arise in traditional sciences. Finally, I argue that Molyneux problems for science will vex automated science especially and that automated learners can solve the Token but not the Type Molyneux Problem.

The department and the PNP program coordinate regular colloquia, typically on Thursdays at 4pm. These represent a major part of the education experience in the department, and graduate students are expected to attend.