all cog sci midterm 1

What is a modular system compared with a centralized one?

Modular: each specialist does one task independently and unsupervised. Centralized: a supervisor assigns and reassigns tasks with constant feedback.

What is encapsulation?

Information from outside a module isn't available inside it.

What is specialization?

A module is restricted to one domain and applies a specific algorithm.

What is localization?

A module is implemented in a circumscribed brain region dedicated to that function.

What are the two extreme views of the mind?

Domain-general (a few general systems using many brain parts) and domain-specific (many task-specific systems in specific brain parts).

What is a top-down effect?

Cognition (beliefs, mood, goals) changing what you perceive.

Why does the Müller-Lyer illusion support encapsulation?

It persists even when you know the lines are equal, so knowledge can't reach into perception.

What is the El Greco fallacy?

If El Greco's vision stretched everything, his canvas would look stretched too, so he'd still paint accurately. A distorted perceiver's reports shouldn't show the distortion.

What is response bias (demand characteristics)?

Participants report what they think the experimenter wants rather than what they perceive.

What is the low-level confound reply to "Black faces look darker"?

Image differences like luminance explain it without categorization. Blurred faces, which can't be categorized, still look darker.

What is the overall verdict on top-down effects in perception?

Even when it looks like our beliefs or knowledge change what we visually perceive, there may be another explanation that happens after perception. (low level

What is a double dissociation?

One patient loses A but keeps B, another loses B but keeps A, suggesting separable systems (prosopagnosia vs. Patient CK for faces vs. objects).

What is the face inversion effect?

Faces are much harder to process upside down than other objects, suggesting specialized face processing.

What is the part-whole effect?

A face part is recognized better within the whole face than alone, suggesting holistic processing.

What evidence supports localization of face processing?

Face-selective neurons in monkeys and face-selective regions in human fMRI (FFA, OFA, fSTS).

What are the two challenges to face specialization?

Expertise (the "Greebles" study) and graded responses of "face cells" to non-faces that share image statistics with faces.

What do all representations minimally have in common?

They are about something, so they can be accurate or inaccurate.

What is the difference between content and format?

Content is what is represented (the meaning). Format is how it is stored (e.g., a picture vs. a sentence).

What is the imagery debate?

A debate over the format of mental imagery: picture-like (Kosslyn) vs. sentence-like (Pylyshyn).

What did the mental rotation experiment show?

Response time rises roughly linearly with rotation angle, which looks like rotating an image.

What is the weakness of the response-time argument for mental images?

Behavior underdetermines format. The same results could come from numerical transformations on sentences.

What is aphantasia?

Little or no voluntary mental imagery, measurable with binocular rivalry, skin conductance, and pupil response.

What is the aphantasia puzzle?

People with aphantasia perform normally on many imagery tasks. They may use another strategy, have unconscious imagery, or be wrong about their own minds.

What is Weber's Law?

Discriminability of magnitudes depends on their ratio, not their absolute difference.

What is synesthesia, and what does it show about format?

One input triggers an experience of another (e.g., letters → color). It shows that content and format can come apart.

What is compositionality?

The meaning of a whole is computed from the meanings of its parts plus their arrangement.

What is the Language of Thought Hypothesis?

The mind uses language-like representations: repeatable elements (names, predicates) recombined by a syntax.

What are systematicity and productivity?

Systematicity: being able to think R(a,b) means being able to think R(b,a). Productivity: thoughts can be chained into arbitrarily larger ones.

What is Dretske's causal theory of content, and its problem?

Content includes all of a representation's causes. Problem: misrepresentation becomes impossible (the disjunction problem).

What is Millikan's selectional theory of content, and its problem?

Content is what the representation was selected to detect. Problem: not all representations were selected, and selection history doesn't clearly pick worm vs. food vs. shape.

What is Dennett's interpretational theory of content, and its problem?

Content is whatever we interpret it to be. Problem: representations aren't causes, and cognitive science stops being objective.

What is the goal of cognitive science?

To explain how the mind works.

What did Hubel & Wiesel (1959) do?

Recorded a single cat visual cortex neuron responding to bars of light at different orientations, showing orientation-selective cells.

What is Marr's bird flight analogy (1982)?

Studying only neurons to understand perception is like studying only feathers to understand flight. You need a higher-level theory (aerodynamics) first.

What are Marr's three levels?

Computational (goal and why), Algorithmic/Representational (representation and algorithm), and Implementational (physical realization).

What does the computational level ask?

What is the goal of the computation, why is it appropriate, and what is the logic of the strategy?

What does the algorithmic level ask?

What is the representation for input and output, and what algorithm transforms it?

What does the implementational level ask?

How are the representation and algorithm physically realized?

What is a representation?

The format in which information is stored or encoded.

What is an algorithm?

A procedure of rules that transforms representations, turning inputs into outputs through intermediate representations.

What is a prediction error?

The mismatch between an expected value and an observed value.

How does the cash register example illustrate the three levels?

Computational: output is the sum of prices. Algorithmic: prices as cents in base 10, summed right to left with carries. Implementational: digits as gear positions.

What did Jonas & Kording (2017) do?

Applied neuroscience analysis methods to a microprocessor running Donkey Kong, asking whether methods that can't explain a chip can explain a brain.

How does Donkey Kong illustrate the three levels?

Computational: the hammer destroys barrels. Algorithmic: has_hammer AND position match. Implementational: position stored in a RAM chip.

How does an AI language model illustrate the three levels?

Computational: predict the next word. Algorithmic: words in context passed through a Transformer. Implementational: runs on GPUs.

What did Huskey et al. (2020) find about personal space?

Computational: distance violations cause negative evaluation. Algorithmic: response = |expected − actual distance|. Implementational: striatum computes the prediction error.

What is the Ultimatum Game, and what did Henrich et al. (2006) find?

Player 1 proposes a split, and Player 2 accepts or rejects (rejection gives both $0). Offers and rejection rates vary widely across cultures.

What are Marr's two lessons?

(1) Distinct, interesting questions exist at each level, so studying the mind isn't just studying the brain. (2) It is very hard to study implementation without higher-level theories first.

What five fields make up cognitive science?

Psychology (what minds do), Neuroscience (how brains work), Philosophy (what minds are), Linguistics (how minds communicate), Computer Science (models what minds do).

What is the computer analogy?

Mental processes are representations operated on by rules.

What is cognitive science contrasted with, and what makes it distinctive?

Phenomenology, psychoanalysis, behaviorism, pharmacology, and biology. It focuses on internal representations, algorithms, and computational-level explanation.

What is the Mirroring View vs. the Constructive View of perception?

Mirroring: the world is passively absorbed like a camera. Constructive: the mind actively reconstructs the world using assumptions. The lecture supports the Constructive View.

What is the Kanizsa Triangle?

An illusory triangle perceived with no triangle drawn (illusory contours), an example of hardwired interpretations.

What is the Ponzo illusion?

Converging perspective lines make same-size objects look different sizes.

What does the shading/bumpiness demo show, and what did Adams et al. (2004) add?

The visual system assumes light comes from above. About 1.5 hours of reversed lighting exposure updates that assumption.

What is the Ames Room?

A distorted room that exploits perspective assumptions so people appear to shrink or grow.

What is the oddball effect?

An unexpected stimulus is perceived as lasting longer, explained by an internal clock that ticks faster for novel stimuli.

What is the flash-lag effect?

A moving object is perceived ahead of a flashed static object. Explanations: forward skew compensating for neural delay, or separate systems with different timing.

What did Tolman (1948) find with rats in mazes?

Rats took novel direct shortcuts, suggesting an internal cognitive map.

What did Tversky & Kahneman (1973) find with the letter K question?

About twice as many words have K as the third letter, but people guess the opposite because words are easier to recall by first letter (availability heuristic).

What did Perl et al. (2023) find about traumatic memories?

For non-traumatic memories, semantic similarity tracks neural similarity. For traumatic memories that link breaks down, suggesting they aren't encoded semantically in the typical way.

What is EMDR, and what mechanism does the lecture propose?

Eye Movement Desensitization and Reprocessing. Overwhelming ("hot") information isn't marked as memory and is re-lived rather than remembered, and reduced attention lets it cool and be stored normally.

What is the Classical Computational Theory of Mind (CCTM)?

The view that the mind literally is a computing system using discrete symbols and rule-based algorithms.

What are the two core claims of CCTM?

Mental representations are discrete symbols, and mental algorithms are sequences of rule-based steps.

What is a symbol in classical computation?

A discrete representation that can be manipulated according to rules.

What is a rulebook in classical computation?

It specifies which operation to perform given the current information or state.

What did Ada Lovelace contribute?

She wrote an algorithm for Babbage's Analytical Engine to compute Bernoulli numbers and recognized that machines could manipulate abstract symbols, not just numbers.

What is ACT-R?

A classical cognitive architecture with declarative memory (symbolic facts linked by relationships) and procedural memory (rules for action).

What is a Turing machine?

A theoretical computer with an infinite tape, a current state, a current position, and a lookup table of rules based on state and the symbol being read.

What is the Church-Turing Thesis?

Any computation a human can carry out with pencil and paper can also be performed by a Turing machine.

What is a Universal Turing Machine?

A Turing machine that takes a description of another Turing machine as input and runs it, making it general-purpose (though sometimes inefficient).

What did McCulloch and Pitts show in 1943?

Logical operations can be implemented by neurons, so neurons can implement Turing-machine computations.

What is the von Neumann architecture?

A computer design that stores programs and data in memory and executes instructions.

How does CCTM challenge Descartes' dualism?

It explains thinking as information processing that can be implemented in purely physical systems.

What is connectionism?

An approach rejecting the claims that all mental representations must be discrete symbols and all algorithms must be explicit step-by-step rules.

What did Descartes (1637) doubt about machines?

That a machine could produce flexible, meaningful language.

What are the strengths and weaknesses of classical, rule-based systems?

Strength: interpretability. Weakness: bad at categorization, generation, extrapolation, open-ended problems, and motor control.

What was the 1950s single-layer perceptron, and what is the XOR problem?

A perceptron draws a straight line separating two classes. XOR patterns can't be separated by any single line.

What happened in the 1980s, 2012, and 2017 in neural networks?

1980s: backpropagation made multilayer training possible. 2012: deep networks surpassed prior systems on ImageNet. 2017: the Transformer was introduced.

What is ImageNet?

A dataset of about 14 million images across 20,000+ categories.

What three factors enabled progress after 2012?

Faster computers, larger networks, and more data.

What are the five basic elements of neural network computation?

Nodes, weights, activation functions, an objective function, and a learning algorithm.

What are the objective function and the loss function?

The objective function defines the correct output. The loss function measures how wrong the current output is (e.g., squared difference, cross-entropy).

What is backpropagation?

It computes how much the loss would change if a parameter changed slightly, using the chain rule.

What is gradient descent?

Updating parameters using the gradient times the learning rate.

What is the general training procedure?

Initialize random weights, input a stimulus, compare the output to the correct answer, adjust weights toward correct, and repeat.

What are a batch, an epoch, and a train-test split?

Batch: a subset of data processed per update. Epoch: one full pass through the training data. Train-test split: separating training data from held-out evaluation data.

What are learning rate annealing and regularization?

Annealing gradually reduces the learning rate over training. Regularization prevents overfitting.

What are the five neural network types mentioned?

Multilayer Perceptron, Convolutional, Recurrent, Diffusion, and Transformer.

What is the key disadvantage of neural networks?

Low interpretability. We often don't understand what they do internally.

How do biological neurons and synapses differ from ANN nodes and weights?

Dendrites perform complex nonlinear computation, synapses are fixed excitatory or inhibitory, and brain connectivity is not fully connected.

Does the brain learn by backpropagation?

The majority view is no, because backprop requires an explicit objective, a loss per input, knowledge of all weights, and precise adjustments.

What are the "maybe yes" pieces of evidence that ANNs resemble human minds?

Zhao & Firestone (2019): humans and networks make similar mistakes. AlphaZero can surface insights that improve human understanding.

What are the "maybe not" pieces of evidence?

Humans need far less training data. Networks can be swayed by irrelevant details (Li et al., 2025) and struggle with intuitive physics (Bi et al., 2025).

What is the "stochastic parrot" critique?

LLMs may be statistical pattern-matchers without genuine understanding.

What are the three alternatives to choosing exclusively between classical and neural computation?

Different levels, hybrid systems, or different phenomena.

What is Othello-GPT?

A neural network trained on about 24 million synthetic games to predict Othello moves.

What are move and position embeddings?

Each is a vector of 512 numbers. Move embeddings represent move locations, and position embeddings represent move order. They are added together and learned during training.

What is an embedding?

A numerical vector representation whose values can encode useful patterns.

What is the residual stream?

The evolving representation produced as information moves through the network.