Lecture 1 + Lecture 1 extra

How can the mind (thoughts, feelings) be created by physical stuff (like the brain)?

Because it was too slow and only studied one tiny thing at a time, failing to see the whole picture.

Cognitive modeling: trying to build a mind on a computer.

To compare the Human Data (what a person did) with the Model Data (what the computer did).

To refine (fix) your theory (your "recipe") if the model's data doesn't match the human's.

To Explain Behavior (e.g., "Why did the player jump?").

To Explain neuronal activation patterns (e.g., "What parts of the brain lit up?").

To Perform Complex Tasks (e.g., build a robot that can win the game).

A model that describes the exact steps the mind takes to get to an answer, not just the final answer.

They help us see why people make mistakes and what the limits of the mind are.

Cognitive fidelity (how well it thinks like a human) and Biological fidelity (how well it copies the physical brain).

Starting with a big theory about the mind (like ACT-R). These have high cognitive fidelity.

Starting by copying the brain cells (like the Blue Brain Project). These have high biological fidelity.

High in cognitive fidelity but low in biological fidelity.

Very high in biological fidelity but low in cognitive fidelity.

A single model that is perfect at both: it acts like a human (high cognitive fidelity) and is built like a real brain (high biological fidelity).

The idea of making the best possible decision given your limitations (bounds).

A limit, like only having two hands (motor bound) or a limited memory (cognitive bound).

Your strategy for doing the best you can given your bounds.

"You did it because it was the 100% perfect, best answer." (Assumes no limits).

"You did it because the environment trained you to."

"You did it because it was the best you could do, given your internal brain limits."

"You did the best you could (your bounds) in that specific situation (the environment)."

You get slow when you try to do two things at once (like press a button and say a word).

It says your brain has one boss (a serial bottleneck) that can only do one thing at a time.

It says your brain has no bottleneck (it can start both tasks in parallel), but it's just hard to coordinate them.

A question about the goal or purpose (e.g., "Why do people text and drive?").

A question about the steps or process (e.g., "What are the steps of texting and driving?").

"What if questions" (e.g., "What if the phone screen was bigger?").

It allows for detailed scrutiny, scientific discussion, making predictions, and "understanding by building."

They require hand-coded rules (brittle to change) and admit a huge space of possible strategies, making it unclear which one users will choose.

Constraints arising from the agent's cognition and body, such as memory limits, perceptual acuity, motor noise, and fatigue.

Constraints arising from the physical world and device design, such as button layout, hardware limitations, and traffic conditions.

Classical architectures use hand-programmed rules, whereas Computational Rationality formulates a POMDP with bounds and uses Reinforcement Learning to *derive* an optimal policy.

They cannot capture adaptation to changing contexts or designs, failing to explain why strategies emerge or differ between individuals.

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

The mental state space (percepts, memory, emotion) plus dynamics and individual capacity limits, which is stochastic and partially observable.

In Computational Rationality, the agent interacts with an *internal* environment (yoked to the external one) where rewards are functions of internal states and bounds are explicitly modeled.

1. Agent solves bounded optimality problems. 2. Internal environment encodes mental states / bounds. 3. External environment includes design/ecology. 4. Interaction is a source of bounds. 5. Preferences are subjective rewards over internal states.

States, Actions, Transition function, Reward function, Observations, and Observation function.

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

He argued that testing isolated verbal hypotheses is slow and fails to build a unified computational theory; we must instead build task-performing models.

How closely the model matches neurobiological details and physical mechanisms of the brain.

How well the model captures complex cognitive functions, algorithmic structure, and behavior.

It mapped "boxes and arrows" to brain regions without specifying the actual computational algorithms or mechanisms performing the tasks.

Models that reveal what information is present in brain activity by trying to predict or classify stimuli based on neural patterns.

Computational theory (what/why), Representation and Algorithm (how), and Implementation (physical realization).

Because the brain is an evolved system where algorithmic solutions are heavily constrained by their neural implementation, so they must be studied together.

Different research groups challenging models using shared tasks and datasets to rigorously test and refine theories linking mind and brain.