September 9
12pm
Carnegie 113
Metacognition is the ability to reason about one’s own thought processes. In this talk, we present a framework for artificial metacognition based on a paradigm of “error detection rules.” We introduce this framework and discuss four results about artificial metacognition, namely (1.) detecting errors is easier than correcting them, (2.) high false-positive error detectors can lead to improved correction for multiple agents, (3.) symbolic representations of cognitive state (a metacognitive output) can themselves provide metacognitive cues, and (4.) data-driven approaches for learning metacognitive cues must be data efficient. We conclude the talk with a preview of our latest work on metacognitive monitoring and correctness that utilizes metacognitive cues derived directly from the vector representations of machine learning output and show how it is robust to adversarial perturbations of model output.
Bio: Paulo Shakarian is the Director of the Syracuse University Institute for Artificial Intelligence where he also holds the K.G. Tan Endowed Professor of Artificial Intelligence and directs the Leibniz Lab. Shakarian has made notable contributions in the areas of logic programming, neurosymbolic AI, security, data mining, and metacognitive AI. He has created the open-source PyReason logic programming platform which has over 145,000 downloads. His academic accomplishments include four best-paper awards, over 100 peer-reviewed articles, 12 issued patents, 8 published books, as well as a startup company that was acquired in 2022. He previously held faculty positions at Arizona State and West Point.