Three Strategies for Salvaging Epistemic Value in Deep Neural Network Modeling

Abstract

Some how-possibly explanations have epistemic value because they are epistemically possible; we cannot rule out their truth. One paradoxical implication of that proposal is that epistemic value may be obtained from mere ignorance. For the less we know, then the more is epistemically possible. This chapter examines a particular class of problematic epistemically possible how-possibly explanations, viz. epistemically opaque how-possibly explanations. Those are how-possibly explanations justified by an epistemically opaque process. How could epistemically opaque how-possibly explanations have epistemic value if they result from a process about which we lack knowledge or understanding? This chapter proposes three different strategies to salvage epistemic value from epistemic opacity, namely salvaging value from 1) functional transparency, 2) modal operator interpretation, and 3) pursuitworthiness. It illustrates using cases from deep neural network modeling.

Philippe Verreault-Julien
Philippe Verreault-Julien
Postdoctoral Researcher

Philosopher working on the ethics, epistemology, governance, and safety of artificial intelligence systems.