Metacognitive Skill Learning in Humans and AI | Stanford

About this Video

• Video Title: Metacognitive Skill Learning in Humans and AI | Stanford
• Channel: Meta-Think
• Speakers: Brandon Conan Smith
• Duration: 00:31:37

Introduction

This Stanford talk by Brandon Conan Smith explores metacognitive skill learning in both humans and AI. The presentation aims to explain metacognitive processes in humans and how this understanding can be applied to AI to enhance its capabilities. The speaker details a formal theory of metacognitive skill learning and discusses its implications for AI design.

Key Takeaways

• Metacognition defined: The ability to monitor and control one's mental states, including attention, emotions, and memory. It's a better predictor of learning and performance than IQ.
• Three types of skills: Perceptual-motor (e.g., tennis), cognitive (e.g., chess), and metacognitive (controlling internal processes like attention). AI currently struggles most with metacognitive skills.
• Dual-system metacognition: Two forms of metacognitive information: fast, automatic feelings (type 1), and slower, declarative strategies (type 2). Type 1 feelings can trigger type 2 strategies.
• Formal theory of metacognitive skill learning: This theory, based on proceduralization (converting declarative knowledge into fast, automatic procedural knowledge), explains how metacognitive skill develops and its hallmarks (power law learning curve). This process is similar to skill acquisition in motor and cognitive domains.
• Applications to AI: The speaker proposes importing the understanding of human metacognitive skill learning into AI design to create more robust, flexible, explainable, and safe AI systems. A potential approach involves a metalearning reward function that automates and refines metacognitive strategies through reinforcement learning.