AI3SD Video: Neural Networks and Explanatory Opacity
Will McNeill · ePrints Soton (University of Southampton) · 2020
Deep artificial neural network (DANN) designers often accept that the systems they construct lack interpretability, are not transparent in other words, that they are inexplicable. It should not be obvious what they mean. Explanations, particularly in the neurosciences, are often thought to consist of the mechanisms which underpin observed phenomena. But DANN designers have complete access to the mechanisms underpinning the systems they build as well as access to their training sets, design parameters, training algorithms and so on. In this talk I distinguish various senses of explanation - ontic, epistemic, objective, subjective. The aims are (1) to help map out the various questions we might be interested in, (2) to scope the limits of mechanistic approaches to the question of explanation, and (3) to try to narrow down the sense in which DANNs are supposed to be explanatorily opaque.