Here is some take-away note about Cognitive Architecture on MIT course 6.S099: Artificial General Intelligence. Original video on YouTube.
Background
Cognitive architecture is a theoretical framework that aims to explain human cognition and intelligence. It is a software instantiation of cognitive theories, which means it is a computer program that implements and tests these theories. The goal of cognitive architecture is to understand and exhibit intelligence across a wide range of tasks and domains. Cognitive architectures specify the aspects of cognition that remain constant across a person’s lifetime, such as attention, perception, memory, and decision-making.
Cognitive Theories
There are a lot of theories and individual findings on carbon brain mechanism. However, we sill don’t have a unified design and working architecture on intelligence. It could be more difficult to synthesize and generalize from these diverse findings without implementation and integration.
The cognitive architectures as software instantiation of cognitive theories was initiated by Allen Newell in 1990. Its goal is to understand and exhibit intelligence across a diverse set of tasks and domains. Cognitive architectures specifies those aspects of cognition that remain constant across the lifetime of an agent, as included:
- Memory systems of agent’s goals and experiences (persistence and self-consistency)
- Knowledge representation
- Learning mechanisms
- Functional processes that lead from perception through to behavior
Active Architectures
- Biological modeling
- Leabra
- SPAUN
- Psychological modelingn
- ACT-R
- CLARION
- EPIC
- Agent functionality
- Companions
- ICARUS
- LIDA
- SIGMA
- Soar
Open issues on CogArch
- Integration of models/agents
- Transfer learning
- Cross-arch comparison
- Multi-model representations, memory, processing
- Related: symbol grounding
- Meta-cognition
- Self-monitoring of agents’ own cognitive processes, goal setting
- Ethical (??)
CogArch vs. Deep Learning
These two tackle different tasks and could be combinational. Deep learning is good at perceptual processing, pattern learning, feature extraction etc. CogArch is suitable for natural encoding known procedures in an associative and reasonable way.
Recommended Reading
- Unified theory of cognition, Allen Newell
- The soar cognitive architecture, John E. Laird
- How to build a brain, A neural architecture for biological cognition, Chris Eliasmith
- How can the human mind occur in the physical universe, John R. Anderson
- Cognitive Architectures: research issues and challenges, Ron Sun etc.
- Cognitive Systems, AI Magazine, Vol. 38 Num. 4, 2017
Recommended Venues
- AAAI, Cognitive Systems Track
- ICCM, Cognitive Modeling
- CogSci, Cognitive Science
- ACS, Advances in Cognitive Systems
- Cognitive Systems Research
- AGI, BICA
- Soar Workshop
- ACT-R Workshop
