Actual Lecture 1 + Lecture 1 Reading

To implement components of cognitive models in computer simulations and evaluate scientific literature on cognitive models.

Type I (Optimality), Type II (Ecological Optimality), Type III (Bounded Optimality), and Type IV (Ecologically Bounded Optimality).

To make psychology more like physics, where a set of universal laws or equations can describe the mind.

It argues that binary yes/no experiments dissect behavior into isolated pieces and fail to explain how underlying laws integrate to produce complex behavior.

Understanding by building (creating process models and simulations).

They show how complex strategy and gameplay emerge from specifying processes for perception and action without a central "brain-to-brain" network.

Complex social behavior can emerge from simple rules and interaction with a human, without needing complex internal emotional parameterization.

Formal specification of theory (working code), detailed scrutiny, prediction of untested conditions, and "understanding by building."

The computational model produces model data, which is compared via statistics to behavioral data produced by the biological organism to refine the theory.

To explain behavior and provide theories of mental processes.

To explain neuronal activation patterns and constrain models to be biologically plausible.

To perform complex tasks and provide computational power/algorithmic tools.

The intersection of Cognitive Science, Neuroscience, and AI, focusing on task-performing, neurally plausible models that explain behavior.

A model that describes the specific information processing steps the mind goes through, typically at Marr's algorithmic level.

Biological fidelity (how well the brain is captured) and Cognitive fidelity (how well the mindehavior is captured).

High cognitive fidelity but lower biological fidelity (top left).

Top right: High cognitive fidelity and high biological fidelity (explaining both mind and brain).

Computational theory (why goal), Algorithmic theory (what steps/ representation), and Implementation theory (how/physical substrate).

The time scale from hundreds of milliseconds to seconds, covering tasks like dialing a phone or choosing a strategy.

Users act to maximize utility/payoff given the limits imposed by their cognitive architecture and the task environment.

External bounds (environment/interface), Internal bounds (cognitive/motor limits), and Policy (strategy).

In classical architectures, strategies are hand-coded rules; in Computational Rationality, the policy is derived via optimization (e.g., Reinforcement Learning).

Partially Observable Markov Decision Process; the formalism used to define the agent's interaction with the environment under uncertainty.

The agent chooses the globally optimal action without any assumed internal bounds (ideal performance).

Behavior is adapted to the statistics of the environment and history (e.g., memory availability reflects usage frequency) rather than internal architectural constraints.

Behavior is optimal given specific internal limits (cognitive architecture) and the immediate task environment (e.g., the PRP bottleneck).

Behavior is optimal relative to both internal agent constraints and broader ecological/environmental constraints (history/context).

Because the behavior (bottleneck/delay) is explained by internal architectural constraints (serial processing) within a specific task, not by environment alone.

Cognitive and physical limits such as memory capacity, peripheral vision quality, motor noise, and the cost of switching gaze.

Constraints from the world, such as road curvature, traffic density, interface layout, and task rules.

Why (computational/optimality reasons), What (algorithmic steps/processes), and What-if (predictions under changed parameters).

Connectivity models, Decoding models, and Representational models.

Decoding models predict stimuli from brain activity; representational models characterize the geometry of the representational space (e.g., RSA).

Because without task-performing computational theories, massive data sets (like connectomes) do not reveal the algorithmic principles of cognition ("20 questions with nature").

Different labs working together to challenge models using shared tasks and datasets to falsify or refine them.

They require hand-coded production rules (making strategy selection unclear) and are brittle (rules must be rewritten for changes in design/context).

Internal bounds (cognition / body limitations) and External bounds (physical world/device design).

Limited peripheral vision, uncertainty about lane position, and noisy motor control.

Phone holder position, traffic density, and road curvature.

In CR, it is derived via optimization (e.g., RL) over a POMDP with bounds; in classical architectures, it is hand-coded by the researcher.

It identifies processing limits and adaptation phenomena to justify the bounds used in models.

It provides the formal framework (POMDP), bounded optimality notions, and algorithms (RL) to compute policies.

They cannot capture how behavior changes with design/context, failing to distinguish between designs or explain individual differences.

Memory (chunking), Perception (eye movements), Motor (speed-accuracy trade-offs), and Environment/Ecology (traffic/workload).

To link behavior to reward maximization under bounds, enabling prediction of adaptation to new designs.

Mental states (percepts, memory, emotion) and their dynamics, including individual capacity limits.

CR POMDPs explicitly model human cognitive bounds (internal environment) and define rewards over internal states, whereas classical POMDPs focus on external environments.

1. Agent solves bounded optimality problems. 2. Internal env encodes mental states/ bounds. 3. External env includes device/ecology. 4. Interaction is a source of bounds. 5. Subjective reward function over internal states.

States (S), Actions (A), Transition function (T), Reward function (R), Observations (Ω), and Observation function (O).

Time (duration), Noise (sensory/motor variability), Uncertainty (imperfect knowledge), and Capacity (limits on resources).

To build task-performing computational models that are both cognitively meaningful and neurobiologically plausible.

Testing isolated verbal hypotheses or mapping regions without building task-performing models fails to reveal mechanistic explanations of cognition.

Connectivity models, Decoding models, and Representational models.

They perform complex tasks with internal representations that can be compared to brain activity (e.g., ventral stream) and behavior.

They cannot be cleanly separated; understanding cognition requires integrating computational, algorithmic, and implementation levels.

A practice where different labs collaborate to challenge and falsify models using shared tasks and datasets.