Why is it difficult to combine individual AI theories to create complex intelligent systems?

Individual AI theories, such as Fitts's Law or the Power Law of Practice, are very task-specific and focus on isolated aspects of human behavior. It's challenging to integrate these distinct theories to understand or create systems for complex tasks like playing chess because there's no clear way to 'smash them together' and get a coherent, reasonable answer or behavior.

What is bounded rationality and how does it influence cognitive architecture design?

Bounded rationality, proposed by Herbert Simon, suggests that humans make decisions rationally but within constraints like problem complexity, cognitive limitations (memory, speed), and available time. Cognitive architecture acknowledges this by designing systems that can 'satisfice'—finding optimal solutions for simplified problems or satisfactory solutions for more realistic ones, rather than always seeking global optimal solutions.

What are the differences between biological, psychological, and functional modeling approaches in cognitive architectures?

Biological modeling (e.g., Spaun) focuses on low-level details like neuronal firing rates and connectivity, aiming for biological fidelity. Psychological modeling (e.g., ACT-R) builds systems reflecting brain areas and interactions to predict human errors and timing. Functional modeling (e.g., Soar, Sigma) prioritizes creating functional systems that solve problems and exhibit cool artifacts, often taking inspiration from human cognition but not strictly adhering to biological or psychological fidelity.

How does ACT-R use MRI data to validate its predictions?

The ACT-R community makes predictions about which brain areas will be active during certain cognitive operations. They then perform MRIs (Magnetic Resonance Imaging) on humans performing those tasks, gather the brain activity data, and correlate it with the system's predictions, providing empirical validation for the architecture's claims about human cognition.

What is the Rosetta project and how does it enable human-robot teaching?

The Rosie project (correctly 'Rosetta' in other references, but 'Rosie' in the transcript) is a multimodal learning system that allows a robotic arm to learn new tasks from natural language and pointing gestures. The robot can understand concepts like colors and object types, and learn action sequences, enabling it to generalize and perform newly taught tasks with different parameters.

How does Soar's memory management deal with forgetting, and why is it beneficial?

Soar implements a 'forgetting' mechanism based on human recency and frequency effects in memory. Memories are decayed and potentially forgotten if they haven't been used recently or frequently, or if they can be approximately reconstructed. This process improves efficiency by preventing short-term memory overflow in tasks like mobile robotics mapping or reinforcement learning state-space management, surprisingly leading to better or comparable performance than systems with no forgetting.

What are some of the open challenges in the field of cognitive architecture?

Key challenges include integrating different cognitive systems and theories, improving transfer learning across diverse tasks, developing multimodal representations that combine symbolic, visual, and auditory information, and enabling metacognition (agents self-assessing their own state and processing). The question of what would happen if human-level AI is achieved also remains a significant societal challenge.

How do cognitive architectures like Soar integrate with modern deep learning methods?

Cognitive architectures can interface with deep learning agents through APIs, using deep learning for sub-problems like object recognition or function approximation. While cognitive architectures typically manipulate discrete symbols, deep learning can process low-level non-symbolic inputs (e.g., pixels) to generate stable symbolic representations, forming a layered system where each approach handles tasks it's best suited for.

Key Moments

MIT AGI: Cognitive Architecture (Nate Derbinsky)

Lex Fridman

Science & Technology4 min read91 min video

Mar 20, 2018|66,458 views|1,320|68

mit agi northeastern artificial intelligence ai cognitive modeling cognitive architecture soar actr cognition psychology cognitive science

Save to Pod

Key Moments

TL;DR

Nate Derbinsky discusses cognitive architectures, a research approach to AGI, focusing on Soar and its application.

Key Insights

Cognitive architecture is an approach to Artificial General Intelligence (AGI) that aims to understand and build human-level intelligence by integrating various cognitive processes.

Historical context shows a progression from individual AI tasks to the need for unified theories, leading to cognitive architectures.

Key cognitive architectures like ACT-R and Soar provide frameworks for modeling human cognition and building intelligent systems.

Soar, a prominent cognitive architecture, emphasizes efficiency, task independence, and is publicly available, with applications ranging from robotics to game playing.

Forgetting, modeled through mechanisms like base-level decay, is crucial for efficient memory management and improved performance in cognitive systems, contrary to human aversion to forgetting.

Integrating modern AI techniques like deep learning with symbolic reasoning in cognitive architectures remains an active area of research, with challenges in grounding and representation.

THE QUEST FOR ARTIFICIAL GENERAL INTELLIGENCE (AGI)

The talk introduces Artificial General Intelligence (AGI) as the pursuit of systems exhibiting human-level intelligence, characterized by persistence, robustness, continuous learning, and the ability to tackle novel tasks. Nate Derbinsky contrasts this with current AI, highlighting the aspirational nature of AGI often depicted in popular culture. He emphasizes that AGI systems should ideally be teachable and adaptable, moving beyond the current limitations of voice assistants like Alexa.

COGNITIVE MODELING AS A PATH TO AGI

The field of cognitive architecture is presented as a multidisciplinary approach to AGI, drawing from neuroscience, psychology, and cognitive science. It involves understanding human cognition at various levels: from acting intelligently (like the Turing Test) to thinking like humans, predicting human behavior, and even understanding the rational rules computers might follow. The core idea is to move beyond isolated theories to a unified framework for intelligence.

THE BIRTH OF UNIFIED THEORIES AND COGNITIVE ARCHITECTURE

Inspired by unified theories of cognition, cognitive architecture seeks to integrate fundamental assumptions about intelligent agents' fixed mechanisms and processes, such as representations, learning, and memory. This integration, when implemented in a system, provides constraints that limit the design space, ultimately guiding progress towards understanding and exhibiting human-level intelligence. This scientific approach is likened to a 'lakatosian research program' where core beliefs evolve over time.

KEY PRINCIPLES AND ARCHITECTURES: ACT-R AND SIGMA

Several core assumptions underpin cognitive architectures, including Newell's time scales of human action, which posit regularities occurring at different temporal levels. Herb Simon's concept of 'bounded rationality' acknowledges human cognitive limitations, leading to 'satisficing' solutions rather than optimal ones. The talk briefly touches upon biological modeling (e.g., Spohn) and psychological modeling (e.g., ACT-R), highlighting ACT-R's focus on predicting human performance and brain activity. Sigma, a newer architecture, aims to unify modern machine learning with cognitive principles.

THE SOAR COGNITIVE ARCHITECTURE: DESIGN AND APPLICATIONS

Soar is detailed as a leading cognitive architecture, emphasizing efficiency and task universality. It operates on a cycle of perception, decision-making, and action, with working memory represented as a directed graph and knowledge encoded in production rules. Soar's efficiency is crucial, aiming for processing under 50 milliseconds. Its applications are diverse, spanning robotics, natural language processing, HCI, simulations, and even games like liar's dice, demonstrating its versatility and broad applicability.

THE BENEFIT OF FORGETTING IN MEMORY SYSTEMS

A key research finding discussed is the beneficial role of forgetting in cognitive systems. By modeling human memory's recency and frequency effects, Soar implements mechanisms like base-level decay to manage memory load. Forgetting non-essential or infrequently used information significantly improves performance and efficiency, especially in large-scale tasks like robotics mapping or reinforcement learning games. This principle of 'forgetting' allows systems to prioritize and reconstruct crucial information, enhancing overall functionality.

INTEGRATION AND FUTURE CHALLENGES IN AGI

Current research in cognitive architectures faces challenges in system integration, transfer learning, and developing multimodal representations that combine symbolic and sub-symbolic processing. Metacognition, or an agent's self-awareness of its own processing, is another area of active development. The ethical implications of succeeding in AGI are also considered, drawing parallels to fictional portrayals of human-like robots. The relationship between cognitive architectures like Soar and deep learning methods is often complementary, with each excelling in different problem domains.

THE DIALECTIC OF RESEARCH: FROM THEORY TO APPLICATION

The research process within the cognitive architecture community, particularly for Soar, involves a cycle of improving the architecture, applying it to solve problems, identifying limitations, and refining the architecture further. This iterative process ensures that the systems remain useful, task-independent, and efficient. The development of mechanisms like forgetting exemplifies how insights from human cognition can lead to practical improvements in AI systems, often with surprising benefits across different application domains.

Mentioned in This Episode

●Software & Apps

●Companies

●Organizations

●Books

●Studies Cited

●Concepts

●People Referenced

Common Questions

Cognitive architecture is a research field that integrates neuroscience, psychology, cognitive science, and AI to understand and build intelligent systems. It aims to develop a core set of fixed mechanisms and processes that intelligent agents use across various tasks, serving as one approach to achieve Artificial General Intelligence (AGI).

Topics

Memory Systems Machine Learning Integration Learning Algorithms Neuroscience & the Brain AI & Machine Learning Cognitive Modeling Artificial General Intelligence Symbolic AI Mindset & Self-Improvement Human Cognition

Mentioned in this video

Concepts

Turing Test

A benchmark for artificial intelligence, proposing that a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.

Power Law of Practice

A psychological law stating that the time to complete a task decreases as a power function of the number of trials.

Newell's Time Scales of Human Action

A framework proposed by Alan Newell outlining different time scales at which human actions and cognitive processes occur, from neuronal to social interactions.

Temporal Difference Learning

A class of model-free reinforcement learning methods which learn by bootstrapping from the current estimate of the value function, with proposed similarities to dopamine processes in the brain.

Fitts's Law

A predictive model in human-computer interaction and ergonomics that quantifies the time required to rapidly move to a target area, based on the distance to the target and its size.

Bounded Rationality

A concept introduced by Herbert Simon, suggesting that human decision-making is rational but subject to limitations in information, cognitive ability, and time.

Physical Symbol System Hypothesis

Proposed by Newell and Simon, it states that symbols and rules are necessary and sufficient for intelligent action, implying that AI can be achieved through symbolic manipulation.

Biologically Inspired Cognitive Architectures

A field of research and associated conference focused on cognitive architectures inspired by biological systems.

People

Alan Newell

One of the founders of AI and a key figure in psychology, who proposed 'Unified Theories of Cognition' which led to the concept of cognitive architectures.

Nate Derbinsky

Professor at Northeastern University working on computational agents that exhibit human-level intelligence, and the speaker of this presentation.

John Anderson

The developer of the ACT-R cognitive architecture and key researcher in the field of cognitive tutoring systems. Also known for his work on rational analysis of memory.

John Laird

A student of Alan Newell, an advisor to Nate Derbinsky, and a co-founder of Soar Technology, known for his work on the chunking mechanism in Soar.

Imre Lakatos

Philosopher of science who proposed the concept of 'Lakatosian science', where progress involves a core set of beliefs surrounded by evolving hypotheses, informing how progress is tracked in cognitive architecture research.

Herbert Simon

Nobel Prize winner in Economics, known for his concept of 'bounded rationality', suggesting that humans operate under cognitive and environmental constraints when making decisions.

Bonnie John

A researcher who applied ACT-R to predict how humans would use computer interfaces while working at IBM, developing tight feedback loops for designers.

Aaron Sloman

The creator of the Spaun model, who wrote the book 'How to Build a Brain' and developed a toolkit for constructing cognitive circuits.

Paul Rosenbloom

The developer of the Sigma cognitive architecture at USC, and one of the prime developers of Soar at Carnegie Mellon.

Chittaranjan Mohan

A researcher at the University of Michigan's Soar group, who worked on the Rosie project, which focuses on learning through text descriptions and multimodal experiences.

Books

How to Build a Brain

A book by Aaron Sloman that details the Spaun model and provides tools for constructing cognitive circuits.

Cognitive Systems Research

An academic journal that publishes a lot of good research on cognitive systems.

How the Human Mind Occurs in the Physical Universe

One of the core ACT-R books, detailing the psychological underpinnings and internal workings of the ACT-R architecture.

Unified Theories of Cognition

A concept proposed by Alan Newell, advocating for a single theory to integrate various specific findings in psychology and AI into a coherent cognitive architecture.

AI Magazine

An academic journal that recently dedicated an entire issue to cognitive systems, making it a good resource for current research.

Advances in Cognitive Systems

A journal and conference series focused on research in advanced cognitive systems.

Cognitive Architecture: A Survey of Metatheoretical Issues

A book that provides a comprehensive survey of cognitive architectures, covering historical context and new features.

Companies

IBM

A technology company that hired Bonnie John to develop software for predicting human-computer interaction based on cognitive models.

Soar Technology

A company founded by John Laird in Ann Arbor, Michigan, which uses Soar among other general intelligence systems, particularly in defense applications.

Software & Apps

Alexa

Amazon's AI assistant, used by the speaker as an example of current AI limitations compared to desired functionality, like teaching the AI new tasks.

Spaun

A large-scale neural model that simulates various cognitive functions, including vision, working memory, and action, explicitly modeling low-level biological details.

Rosie

A system mentioned as working towards teaching AI new tasks, moving beyond simple command-response, using multimodal inputs.

Sigma

A relative newcomer cognitive architecture developed by Paul Rosenbloom at USC, capable of modern machine learning, vision, and optimization using factor graphs and message passing.

Michigan Liar's Dice

A free iOS App Store game developed using Soar, allowing users to play Liar's Dice against a Soar agent with adjustable difficulty.

Soar

A cognitive architecture focused on efficiency and real-time processing, developed by Alan Newell's students, widely used in various applications from robotics to games.

ACT-R

A cognitive architecture that models human cognition by integrating procedural and declarative memory systems, used for psychological modeling and predicting human behavior.

TensorFlow

Google's open-source machine learning framework, mentioned as the appropriate tool for tasks like object detection, contrasting with the applications of cognitive architectures like Soar.

Lisp

A programming language in which the main distribution of ACT-R is implemented.

Erlang

A programming language used by the Air Force Research Lab to implement ACT-R for parallel processing.

SWIG

A software development tool that Soar uses to generate bindings, allowing it to interface with different programming languages and platforms.

Organizations

Northeastern University

The institution where Nate Derbinsky is a professor, conducting research on intelligent computational agents.

Air Force Research Lab

A research laboratory in Dayton that implemented ACT-R in Erlang for parallel processing of large declarative knowledge bases.

University of Michigan

The institution where the Rosie project is being developed, and where Nate Derbinsky made donations.

BBC

The broadcasting company that produced the television series 'Humans', which explores scenarios of human-level AI.

University of Southern California

The institution where Paul Rosenbloom developed the Sigma cognitive architecture.

Georgia Institute of Technology

An institution where a Soar-based interactive art installation called the 'Dom' was created.

Carnegie Mellon University

An institution where the ACT-R cognitive architecture was developed and is actively being researched.

National Research Laboratory

An organization that has a Java port of ACT-R which they use in robotics.

Institute for Creative Technology

An institute associated with USC where Sigma is becoming the basis for the Virtual Human Project.

Media

Knight Rider

A TV show that inspired the speaker's interest in AI, featuring an intelligent car named KITT.

Humans

A BBC television series recommended by the speaker for its exploration of the societal implications of human-level AI and 'synths' (robots that look and interact like humans).

Pirates of the Caribbean

A film series where a scene depicting the game 'Liar's Dice' serves as a real-world example of the game implemented in a Soar system.

Studies & Research

Rational Analysis of Memory

A research approach, pioneered by John Anderson, that examines human memory by analyzing patterns in large text corpora to derive models of how memory should ideally function for a given environment.