Cognitive Architectures

What Makes Something a Cognitive Architecture

A cognitive architecture is more than a program that solves a specific task. It is a general framework that can, in principle, support any cognitive task by combining a fixed set of computational mechanisms. The analogy is to computer hardware architecture: just as a CPU has fixed components (registers, arithmetic logic unit, cache) that can be combined to execute any program, a cognitive architecture has fixed cognitive components (working memory, long-term memory, production rules, perceptual buffers) that can be combined to perform any cognitive task.

To qualify as a cognitive architecture, a system must typically satisfy several criteria. It must have a theory of memory, specifying how knowledge is stored, retrieved, and updated. It must have a theory of processing, specifying how information flows between components and how decisions are made. It must have a theory of learning, specifying how the system acquires new knowledge and skills from experience. And it must be general-purpose, capable of performing a wide range of cognitive tasks rather than being specialized for a single domain.

The field has produced over 300 distinct cognitive architectures over the past several decades, though only a handful have been developed to the point of supporting extensive empirical research. The most influential are ACT-R and Soar, which between them account for the majority of published research in computational cognitive modeling.

ACT-R: The Modular Mind

ACT-R (Adaptive Control of Thought, Rational), developed by John Anderson and colleagues at Carnegie Mellon University since the 1970s, is the most widely used cognitive architecture in psychology and cognitive science. It models the mind as a collection of independent modules, each responsible for a specific type of information processing, communicating through a central production system.

The core modules of ACT-R include: a declarative memory module that stores facts and experiences as chunks (structured knowledge units), a procedural memory module that stores production rules (if-then rules that specify actions to take when specific conditions are met), a visual module that processes visual input and maintains a representation of the currently attended visual object, an aural module for auditory input, a manual module for motor output, and a goal module that maintains the current task context.

These modules operate in parallel, each processing its own type of information simultaneously. However, they communicate through the production system, which can only fire one production rule per cycle (roughly 50 milliseconds, matching the observed timing of human cognitive operations). This serial bottleneck is a deliberate design choice that corresponds to the observed serial nature of human conscious attention and predicts specific patterns of dual-task interference that match experimental data.

ACT-R memory retrieval is governed by a mathematical equation that calculates the activation level of each memory chunk based on its recency and frequency of use, its relevance to the current context, and a spreading activation mechanism that boosts chunks related to the current focus of attention. This activation equation predicts specific patterns of memory retrieval time, accuracy, and errors that match human performance across hundreds of experimental tasks, from simple memory recall to complex problem solving.

A key strength of ACT-R is its integration with brain imaging data. Each ACT-R module has been mapped to a specific brain region (the declarative module to the hippocampus and temporal cortex, the goal module to anterior cingulate cortex, the procedural module to basal ganglia), and the predictions of the model match patterns of brain activity observed in fMRI studies of the same cognitive tasks.

Soar: Problem Solving and Learning

Soar, developed by John Laird, Allen Newell, and Paul Rosenbloom at the University of Michigan since the 1980s, organizes cognition around a cycle of proposing, evaluating, and applying operators that transform the current problem state toward a goal state. Where ACT-R emphasizes the modularity of cognitive processing, Soar emphasizes the universality of the problem-solving process.

In Soar, all cognitive activity is framed as problem solving. Perceiving the world means constructing an internal representation (the problem state). Deciding what to do means selecting an operator to apply to that state. Executing the decision means applying the operator, which transforms the state. Learning means acquiring new knowledge about which operators are effective in which situations.

The distinctive mechanism of Soar is its approach to learning through impasses. When the system cannot decide which operator to apply (a decision impasse), cannot predict the result of applying an operator (a state no-change impasse), or finds that the current state already satisfies the goal (a state no-change impasse), Soar creates a subgoal and recursively applies the same problem-solving process to resolve the impasse. When the impasse is resolved, the result is compiled into a new rule through a process called chunking, so that similar situations in the future can be handled without creating a subgoal.

Modern versions of Soar have incorporated working memory, semantic memory, and episodic memory as distinct systems, bringing it closer to ACT-R modular approach while retaining its distinctive emphasis on problem solving and learning. Soar has been applied to complex real-world tasks including air combat simulation, robot navigation, instructable agents, and game playing.

Newer and Hybrid Architectures

Several newer cognitive architectures attempt to bridge the gap between the symbolic processing of traditional architectures and the neural processing of connectionist models.

LIDA (Learning Intelligent Distribution Agent), developed by Stan Franklin at the University of Memphis, implements Global Workspace Theory as a computational architecture. Global Workspace Theory proposes that conscious processing involves the broadcast of information from a central workspace to multiple specialized processors, and LIDA implements this through a cognitive cycle in which perceptual information is processed by specialized "codelets," relevant information competes for access to the global workspace, and the winning information is broadcast to all modules for further processing. LIDA is particularly interesting for artificial brain research because it provides a computational account of consciousness that can be directly implemented and tested.

Sigma, developed by Paul Rosenbloom at USC, uses probabilistic graphical models as a unifying mathematical framework. In Sigma, all knowledge (declarative facts, procedural rules, perceptual inputs, motor outputs) is represented as factors in a graphical model, and all cognitive processing (memory retrieval, decision making, learning) is implemented as probabilistic inference over this model. This gives Sigma a unified mathematical foundation that naturally supports both symbolic and subsymbolic processing.

CogNGen combines vector-symbolic memory representations with deep neural network components, creating an architecture that can learn from raw sensory data (like a neural network) while also supporting the structured knowledge manipulation (like a cognitive architecture) needed for reasoning and planning.

Relevance to Artificial Brain Research

Cognitive architectures are important to artificial brain research for several reasons. They provide empirically validated models of human cognitive performance that any artificial brain should be able to match. They identify the functional components that a mind requires, independent of the neural implementation. And they suggest organizational principles (modularity, serial bottleneck, multiple memory systems) that may be essential for achieving flexible, general-purpose intelligence regardless of the underlying computational substrate.

The relationship between cognitive architectures and neural models is increasingly seen as complementary rather than competitive. A cognitive architecture specifies what the mind does; a neural model specifies how the brain implements it. Building an artificial brain may ultimately require both levels of description, cognitive architectures to ensure the system has the right functional organization and neural models to ensure it operates with the efficiency and adaptability of biological neural tissue.