AI+Human Objectives Initiative: What does it mean to implement an algorithm, and why does it matter?

Abstract: Much of cognitive science is premised on the idea that mental processes involve computation. A difficult foundational question is: what does it take for a physical system (e.g., a brain) to implement a computation or algorithm? We will explore some possible answers to this question, guided in part by tools from the theory of causality and in part by new experiments with artificial neural networks. In many ways we are currently in a better position than we were several decades ago to answer these questions, and also to appreciate some subtleties that cause problems for most previous philosophical takes on the topic. Possible implications for AI interpretability and safety will also be discussed.