Compositional learning in humans vs AI
Abstract: Compositionality---the way people combine concepts to form more complex concepts---has often been identified as a point of difficulty for AI. In this talk I'll present results comparing human and machine performance in on a common platform, the "Game of Hidden Rules" (GOHR). The GOHR is a simple rule-discovery game in which a player (human or AI) tries to classify objects into categories based on an unknown rule that they must infer by trial and error. In a series of experiments, we compared human and AI (RL) performance on "compositional" rule sequences, in which later rules were constructed from combinations of earlier ones. For example, if the learner first solves a rule involving a color match, and then one involving shape, does that make it easier to solve a later rule involving both color and shape? That is, does simpler learning "transfer" to composite learning? The results show that human and AI simple-to-complex transfer are almost completely uncorrelated, suggesting that contemporary AI does not yet effectively reflect the way that humans learn from prior experience.
Bio: Dr. Jacob Feldman, Professor of Psychology and Cognitive Science, Rutgers University Self-Reference and Metacognition Human and Artificial Intelligence
Self-Reference and Metacognition Human and Artificial Intelligence
Abstract: Our capacity for self-reference allows us to introspect on our thoughts and preferences, engage in metacognition, and evaluate our progress on projects while reconfiguring our approach as needed. Despite being central to human intelligence, it has received marginal attention in developing artificial intelligence. There are burgeoning signs of an artificial equivalent of our capacity for self-reference. To illustrate, so-called self-attention mechanisms in Large Language Models, meta-reinforcement learning modules, and second-guessing can each be argued to be examples of artificial self-reference. However, in each case, the capacity is rigid in scope compared to our human capacity, and it developed as a side effect of pursuing other goals. To date, no explicit attention has been directed at artificially replicating this central element of human intelligence. The paper argues that a general capacity for self-reference is key to bringing AI to the next level.
Bio: Dr. Susanna Schellenberg, Distinguished Professor of Philosophy and Cognitive Science, Rutgers University