MIT AGI: Cognitive Architecture (Nate Derbinsky)

The speaker, Nadir Bin Ski, introduces the concept of cognitive architecture as an approach to achieve Artificial General Intelligence (AGI). He discusses the importance of understanding human-level intelligence and the role of cognitive science in AI development. Bin Ski shares his inspiration from TV and movie characters and outlines the historical context of cognitive architecture, highlighting its interdisciplinary nature.

Bin Ski delves into the intricacies of cognitive modeling, emphasizing the need to understand and predict human behavior at multiple levels. He contrasts cognitive modeling with other AI systems, highlighting the importance of robustness, learning over time, and the ability to handle tasks unknown beforehand. The speaker also touches upon the categories of individuals in the field, from curious learners to systems developers, and the significance of transfer learning.

The speaker references Alan Newell's work on the unified theories of cognition, which proposes a core set of mechanisms and processes that intelligent agents use across tasks. Bin Ski discusses the implementation of these theories in cognitive architectures and the challenges of integrating individual theories to build intelligent systems. He also explores the concept of lactose ium, a form of science that involves testing and refining hypotheses through iterative processes.

Bin Ski introduces Newell's time scales of human action, which categorizes cognitive processes based on the time they operate. He explains the concept of bounded rationality, which considers the limitations under which humans make decisions. The speaker also discusses the simple system hypothesis and the idea that intelligent systems can be implemented in silicon-based computers, not just in biological systems.

The speaker discusses the different levels of modeling in AI, from biological and psychological to sociological and economic. He highlights the work of Lieber and Spawn in modeling low-level neuronal details and the efforts to build tasks above this layer. Bin Ski also mentions the focus on functional systems that solve complex problems, such as Soar and Sigma, and their applications in various fields.

Bin Ski provides an overview of a prototypical cognitive architecture, detailing its components such as perception, short-term memory, knowledge representation, and action selection. He explains the cyclical process of perception, knowledge application, and action in response to the environment. The speaker also touches upon the importance of reactivity and the ability to learn and adapt over time.

The speaker focuses on the Soar cognitive architecture, discussing its development, applications, and unique features. He mentions the work of John Laird and Paul Rosenbloom, who were students of Alan Newell, and their contributions to Soar. Bin Ski highlights Soar's focus on efficiency, its public distribution, and its ability to run on various platforms. He also shares examples of Soar's applications in diverse areas, including computer hardware selection, natural language processing, and robotics.

Bin Ski discusses the use of Soar in creative and interactive applications, such as the 'Lumina' project at Georgia Tech, which explores the relationship between humans and machines through a dance installation. He also mentions the 'Michigan Liar's Dice' game, which showcases Soar's ability to learn and adapt in a gaming context. The speaker emphasizes the diversity of Soar's applications and its potential for collaboration and co-creation with humans.

The speaker explores the concepts of learning and forgetting in cognitive systems, using Soar as a case study. He discusses the implementation of a forgetting mechanism based on human memory properties, such as recency and frequency effects. Bin Ski presents the surprising finding that forgetting can improve system performance, as seen in applications like mobile robotics and dice games. He also touches upon the challenges of integrating different types of memory and the importance of meta-cognition in AI systems.

Bin Ski continues the discussion on forgetting, explaining its role in improving the efficiency and performance of cognitive systems. He provides examples of how forgetting can be beneficial, especially in tasks with large state spaces. The speaker also addresses the question of whether forgetting is essential for AGI, suggesting that it may be necessary for systems to model human-like behavior and interact effectively with humans.

The speaker addresses the question of integrating deep learning with cognitive architecture, such as Soar. He discusses the potential for combining the two approaches and the challenges involved, particularly in terms of grounding and representing sensory data in a symbolic way. Bin Ski mentions the work of other cognitive architectures like Spawn and Sigma in dealing with these issues and suggests that there is room for further exploration and development in this area.

Bin Ski concludes by highlighting some open questions and future directions in the field of cognitive architecture. He mentions the challenges of integrating systems over time, transfer learning, and the development of multimodal representations. The speaker also discusses the importance of metacognition and the ethical considerations of creating human-level AI. He provides recommendations for further reading and venues for staying informed about developments in cognitive systems.