The Cognitive Science Master Study Guide

The Prehistory of Cognitive Science

Behaviourism dominated psychology in the early 20th century, restricting it to observable behaviour. Cracks in its account of conditioning, the rise of computational models of mind, and discoveries about attention pushed psychology toward cognitivism.

1.1 Two assumptions of behaviourism

1. All learning is the result of conditioning.
2. Conditioning depends on association and reinforcement.

The behaviourist slogan: "Psychology is the science of behavior." Mental processes are unobservable and therefore unscientific. Chomsky's killer rejoinder:

"Defining psychology as the science of behavior was like defining physics as the science of meter reading."— Noam Chomsky

1.2 Classical (Pavlovian) conditioning

Pavlov initially studied salivation in dogs (Nobel Prize 1904). When he noticed dogs salivating to the assistant opening the door — a "psychic secretion" — he pivoted his entire research programme.

The big debate inside classical conditioning

Rescorla (1973) tested this with rats. Pair light + loud sound → both elicit freezing. Then habituate half the rats to the sound alone until they stop freezing to it. Now present the light:

• S–R predicts: freezing to the light is intact (the light → freezing bond is independent).
• S–S predicts: no freezing — the light triggers the representation of the now-habituated sound.

Result: habituated rats did NOT freeze to the light — strong support for the cognitivist S–S theory.

Phenomena to know

• Extinction — CR weakens when CS is presented repeatedly without the US.
• Spontaneous recovery — after extinction, the CR partially returns after a rest. Pavlov concluded the CR is inhibited, not lost.
• Generalization — CR also occurs to similar stimuli (tones near the trained tone).
• Discrimination — organism learns to respond only to a specific CS, not similar ones.
• Volkova (1953) — semantic generalization: children conditioned to "good"/"bad" generalized to whole sentences like "The children are playing nicely together" / "The Fascists destroyed many cities."

📘 Textbook adds — Tolman's latent learning & cognitive maps (1930, 1946)

📘 Textbook adds — Lashley (1951): "The Problem of Serial Order in Behavior"

Lashley argued that complex behaviour (speech, tennis, piano playing) cannot be a chained sequence of stimulus–response links because what happens next depends on what will happen later in the sequence and on the overall goal. He proposed behaviour is organised hierarchically, with high-level plans broken down into sub-plans. Two foundational ideas crystallise from his essay:

• Subconscious information processing — most of the planning that turns goals into movements happens below awareness.
• Task analysis — a complex cognitive ability can be understood by decomposing it into a hierarchy of simpler sub-tasks (the methodological backbone of cognitive science).

1.3 Operant (instrumental) conditioning

Reinforcement schedules

Ratio > Interval because in ratio schedules reinforcers scale with response rate; in interval schedules they're time-capped.

Positive reinforcement = arrival of a stimulus increases the response. Negative reinforcement = removal of a stimulus increases the response. Both make behaviour more likely (≠ punishment).

Shaping, discrimination, concept learning

• Shaping = training complex behaviour by reinforcing successive approximations (dog → kitchen → refrigerator → door → scratching).
• Discriminative stimulus = signal that a response will be reinforced (a light being on).
• Concept learning — pigeons rewarded for pecking Monet (not Picasso) generalize to Cézanne, Renoir → a category "impressionist" forms.

1.4 Cognition and computation

If humans can simulate a single-tape Turing machine (slowly, inefficiently), then the brain is Turing-complete. McCulloch & Pitts built networks of neurons from three principles:

1. Basic physiology
2. Propositional logic
3. Turing's theory of computation

Their results: any computable function can be computed by a network of neurons; all logical operators can be built from simple neural networks.

📘 Textbook adds — Chomsky's Syntactic Structures (1957)

Chomsky distinguished the deep structure of a sentence (its constituent phrase structure) from its surface structure (the actual word order, derived via transformational rules).

This was the first time a linguist offered an explanatory account of language structure rather than just classification — the model for algorithmic theories of mental capacities.

1.5 The mind as an information processor

George A. Miller (1956) — "The magical number 7 (± 2)": human channel capacity ≈ 3 bits ≈ 7 items, roughly independent of modality. Measured by:

• Digit-span task — repeat back the longest sequence of digits you can hold.
• Absolute judgment task — identify stimuli along one dimension.

Psychophysics — Weber's law

• k = 0.03 for weight (3% change detectable)
• k = 0.01 for length
• k = 0.25 for sound frequency in mice

The same absolute difference (10 units) is easy to detect at low magnitude (10 vs 20) and hard at high magnitude (110 vs 120). The detection probability curve runs sigmoidally from 0% (no difference) through 50% (the jnd) to 100% (clear difference).

1.6 Attention — reducing information load

Cherry's dichotic listening / shadowing: participant repeats one ear's story aloud and cannot report the content of the other ear. They do notice physical changes (voice pitch shift, sudden tones).

Three arguments AGAINST early selection

Three alternatives to early selection

Three Milestones of Cognitive Science

Three foundational achievements: SHRDLU (language as algorithmic processing), the imagery debate (spatial vs propositional representation), and Marr's three levels (computational / algorithmic / implementational). Cognitive science matures by treating the mind as a system that operates over internal representations.

2.1 Language and micro-worlds

SHRDLU's three processing stages:

Ambiguity demonstration: "Put the red cube on the block in the box." Two readings — [red cube on the block] / [in the box] vs [red cube] / [on the block in the box]. Syntactic parsing alone is insufficient; world knowledge is required to disambiguate.

2.2 The imagery debate — what is mental imagery?

Evidence favouring the spatial view:

• Detail recognition — to identify small details in an imagined object, you must "zoom in" — incompatible with propositional format.
• Physical-property effects — brightness, contrast, motion speed affect reaction times the same way for perceived and imagined stimuli.
• "Imagine your dinner" effect — people with bigger houses take longer to mentally answer spatial questions about them. It takes time to travel in our mind.
• Mental rotation (Cooper & Shepard 1973) — reaction time to judge whether the letter R is normal or mirrored grows linearly with the angle of rotation, peaking near 180°.

2.3 Marr's three levels of analysis

David Marr (1945–1980), neuroscientist of vision. Two prizes are named after him (IEEE/ICCV; Cognitive Science Society).

Exam trap: classic distractors swap the algorithmic and implementational levels. The algorithm describes representations and steps; the implementation describes the physical substrate.

2.4 Marr's three stages of vision

The 2½-D sketch is viewer-centered (depends on where you stand); the 3-D sketch is object-centered (invariant to viewpoint). The 2½-D stage uses stereopsis and Gestalt laws.

Contour-related demonstrations: Hidden Dalmatian (low contour image, hard to perceive) · Kanizsa's triangle (illusory contours filled in) · Camouflage = reducing contours to defeat the primal sketch.

2.5 Perceptual constancies

2.6 Object categorization — four theories

2.7 Categorization hierarchies — expertise effect

Non-experts default to the basic level; experts categorize at a more specific level within their domain. Example: Barcelona Chair (Mies van der Rohe & Lilly Reich, 1929).

The Turn to the Brain

From the immaterial soul to localized neural circuits. Phrenology is wrong in particulars but not absurd in spirit — some abilities really do map onto specific brain regions. The lecture surveys dualism vs materialism, basic neuroanatomy, and a parade of neuropsychological case studies that establish localization of function.

3.1 Philosophical roots

3.2 Psychiatry then and now

• 1247 — Bethlem Royal Hospital ("bedlam") founded. Closed institution; no treatment, just isolation.
• 18th century — rest, cleanliness, regularity. King George III of England (1738–1820) went mad and recovered — proof that mental illness can pass.
• 20th century — frontal lobotomy: removal/disconnection of the frontal lobe "to reduce the complexity of psychic life." Required a neurosurgeon.
• Walter Freeman (1945) — transorbital lobotomy with an icepick through the eye socket. > 40,000 people in the US. Side effects: cognitive impairment, flattened affect.
• Paul Broca (1824–1880) — founder of clinical neuropsychology. Showed that damage to Broca's area impairs speech but not other abilities — early empirical case for localization.

3.3 The brain in numbers (Carl Sagan's "very big place in a very small space")

• ≈ 1,300 g total mass
• ~ 10¹¹ neurons total (20 × 10⁹ neocortical)
• ~ 15 × 10¹³ cortical synapses
• Two hemispheres joined by the corpus callosum; lateralized despite sensory/motor symmetry.

The neuron

Neurotransmitters

3.4 Aphasias — language disorders

Aphasia affects ~35% of stroke patients.

3.5 Split-brain — corpus callosotomy

Cutting the corpus callosum treats refractory epilepsy. No major IQ, conversational, or coordination deficit. But experiments reveal striking dissociations:

• Object presented to left visual field (→ right hemisphere) — patient says "There is no object."
• But the left hand (also right-hemisphere) can pick out the same object from a collection perfectly.
• Patients confabulate to explain their left-hand's behaviour.

3.6 Catalogue of deficits

3.7 Two visual streams

📘 Textbook adds — Petersen et al. (1988): PET subtraction logic

An early functional-neuroimaging landmark for single-word processing. Subjects performed a hierarchy of four tasks; each scan was subtracted from the one above to isolate the new component:

The pattern of activations supported a parallel (not strictly serial) model of single-word processing. Significance: this study established the paired-subtraction paradigm that all later fMRI cognitive subtraction designs inherited.

📘 Textbook adds — Logothetis (2001): what does BOLD actually measure?

The crucial follow-up question after the fMRI revolution: BOLD measures blood oxygenation — but is that correlated with neuronal output (spikes) or input (synaptic activity)? Logothetis put both fMRI and non-magnetic microelectrodes in an anaesthetised monkey's V1 during a rotating-checkerboard stimulus.

Implication: fMRI activation in a region reflects information arriving there, not necessarily spikes leaving. A region can show BOLD without firing more — undermines naive "this region does X" inferences.

3.8 Capgras delusion — the inverse of prosopagnosia

Delusion that a loved one has been replaced by an identical-looking impostor. Explicit recognition is intact, but the implicit/emotional response is missing. Confabulation: "This person looks like my partner, but I don't feel the same about them — so it must be someone else."

Strategies for Brain Mapping

No single neuroscientific technique sees the whole picture. Each method trades off temporal resolution against spatial resolution. Mastery means knowing which tool to reach for, what it measures, and the conceptual logic of subtraction and double dissociation.

4.1 Anatomical classification

• Surface classification — gross anatomy (gyri, sulci, lobes).
• Cellular classification (Brodmann) — Brodmann used staining to identify ~52 areas with distinct neuronal populations. Principle of Segregation: "Cerebral cortex can be classified into different areas with unique neuronal populations."

DTI (Diffusion Tractography) = MRI-based visualization of white-matter fibre tracts — measures anatomical connectivity, not function.

4.2 The big methodological tradeoff

Exam-critical: fMRI does NOT measure neural activity directly. It measures the BOLD signal: "changes in magnetic properties of haemoglobin in the blood due to brain activation."

4.3 ERP components

An ERP (event-related potential) is a time-locked average of EEG to a specific event.

4.4 The Locus of Selection problem

Does attention modulate processing before or after a stimulus representation is built?

• ERP timing: P1 and N1 are enhanced for attended stimuli — early modulation.
• Macaque microelectrode recordings localize attentional modulation to V1–V4, before object recognition or access to meaning.
• Resolution: combining ERP (timing) with microelectrodes (location) shows attention acts early, in pre-representational visual cortex.

4.5 fMRI logic — subtraction & hierarchical design

To localize a function, contrast a task that requires it against a near-identical task that doesn't. Hierarchical lexical-access design:

Each contrast isolates the new component recruited at that step.

4.6 Owen et al. (2006) — detecting awareness in the vegetative state

Published in Science 313:1402. A patient diagnosed UWS was asked to imagine either playing tennis (engages motor/SMA) or walking through her house (engages parahippocampal regions). She produced the appropriate, task-specific BOLD activation on command — proving covert awareness despite the absence of behavioural response. This paradigm became a yes/no communication channel for "vegetative" patients.

Connectionism

Two starting points in Marr's hierarchy give rise to two paradigms. Symbolic AI starts from the mind (algorithms, interpretable rules). Connectionism starts from the brain (biology has already produced intelligence). The result: networks whose knowledge is a pattern of weights, not a list of beliefs.

5.1 Two starting points

📘 Textbook adds — The Physical Symbol System Hypothesis (Newell & Simon, 1976)

The lecture's "Symbolic AI" pole is grounded in a specific thesis the textbook treats as foundational. Allen Newell & Herbert Simon (Turing Award lecture, 1976):

"A physical symbol system has the necessary and sufficient means for general intelligent action."— Newell & Simon, 1976

Two claims packed in: (i) necessity — anything intelligent must be a physical symbol system; (ii) sufficiency — building one is enough to produce intelligence.

Four defining features of a physical symbol system

1. Symbols are physical patterns (inscriptions on a tape, voltage states, neural firings).
2. Symbols can be combined into complex structures via recursive rules (like sentences in propositional logic).
3. The system contains processes that transform symbol structures in rule-governed ways — this is thinking.
4. Those transformation processes can themselves be represented as symbols within the system (meta-representation).

Cognition, on this view, is heuristic search through a problem space. Newell & Simon's General Problem Solver (GPS) applied means–end analysis: compute the difference between the current and goal state, pick an operator that reduces it, apply, repeat. The PSSH defines what connectionism rejects.

5.2 From biology to schematic neurons

5.3 Learning — the delta rule (single-layer)

The perceptron convergence rule: training will find a solution in every case where a solution is possible. But which functions ARE possible?

5.4 The XOR problem

The single-layer perceptron cannot learn XOR — it oscillates, never converges. Why?

Perceptrons only learn linearly separable functions. XOR is not linearly separable — you cannot draw a single straight line that separates the (1,0) and (0,1) cases from (0,0) and (1,1).

5.5 The escape route — multi-layer networks & backpropagation

• Universal approximation theorem: a multi-layer network can compute any Turing-computable function.
• But the perceptron rule no longer works — hidden units have no target activation.
• Backpropagation calculates each hidden unit's "share of responsibility" for the output error and uses it to update weights.
• Gradient descent: follow the negative gradient of the error surface; stop when the gradient is zero. Risk: local minima ≠ global minimum.

5.6 Biological plausibility — the critiques

• Schematic neurons ≠ real ones; questions of parallelism and scale.
• No evidence backpropagation occurs in the brain.
• How would the brain set the number of hidden units?
• No evidence individual neurons receive error signals from all downstream neurons.
• Most biological learning is not supervised.

5.7 Cognitive implications — distributed vs local representations

• Knowledge lies in a pattern of weights, not in any one unit.
• A trained network does not need a separate unit per feature.
• Processing = input vector × weight vector. No discrete beliefs, no explicit rules.
• Algorithmic in a limited sense: the learning rule and activation function are algorithms — but they're not task-specific and they don't operate over explicit representations.

Conclusion: "The nature of representations and computation in neural networks is fundamentally different compared to physical symbol systems."

Modularity of Mind & Dynamical Systems

Three rival pictures of mental architecture: (1) Fodor's classical modularity — domain-specific input modules plus a non-modular central system; (2) massive modularity (Cisek-style evolutionary view) — no central processor at all; (3) dynamical systems theory — cognition as a process that evolves in time, possibly without representations or computation.

6.1 Agents: three tiers

6.2 Classical (Fodorian) modularity

Aristotelian roots: horizontal faculties (perception, attention, memory) are domain-general; vertical faculties are domain-specific (colour, shape, face/voice, grammar, conspecific recognition).

Evidence cited: lesion studies, Broca's vs Wernicke's aphasia, brain mapping.

6.3 Massive modularity (Cisek 2019)

The radical alternative: there is no domain-general central processor. The mind is hundreds or thousands of genetically specified Darwinian modules selected for specific adaptive problems.

• "The most important thing about the brain is that it evolved."
• Domain-general learning mechanisms cannot detect statistically recurrent domain-specific structure.
• Each module exploits specialized, domain-specific rules.
• Descriptive vs pragmatic representations: control loops only need action-oriented (pragmatic) representations, not world models.
• Input–output functionalism ignores the cyclical nature of behaviour.
• No single decision-making system — just domain-specific competition mechanisms.
• Conceptual maps emerge from learning on top of sensorimotor loops — no symbol-grounding problem.

📘 Textbook adds — The cheater-detection module: Wason & Cosmides/Tooby

The textbook's flagship case study for a Darwinian module. The Wason selection task: four cards (E, K, 4, 7); rule "If a card has a vowel on one side, then it has an even number on the other". Which to turn? Correct answer: E and 7 (modus tollens). Most subjects say E and 4 — a famous failure of abstract conditional reasoning.

Griggs & Cox (1982) reframed the same logical task as a deontic conditional: "If a person is drinking beer, then that person must be over 19" with cards BEER, COKE, 16, 25. Now subjects answer correctly (BEER, 16) at near-ceiling rates.

Cosmides & Tooby argued the improvement reveals a domain-specific, evolved cheater-detection module for social-exchange reasoning. The argument links to the evolution of cooperation via the TIT FOR TAT strategy in indefinitely-iterated prisoner's dilemmas: applying TIT FOR TAT requires identifying defectors, so natural selection would favour a module specialised for spotting them.

6.4 Dynamical systems theory (van Gelder 1995)

"What might cognition be, if not computation?"— Tim van Gelder, 1995

Cognition as a process that evolves through time, not necessarily involving computation or representations.

• State space = geometric space of all possible system states. Each independently varying quantity = one dimension.
• Trajectory = path through state space from initial conditions.
• Two senses of "dynamical system": trivial (anything that evolves in time) vs technical (analyzable with DST tools).

📘 Textbook adds — ACT-R as a hybrid architecture (Anderson, CMU)

Where Soar (Newell, Laird, Rosenbloom) is purely symbolic, ACT-R ("Adaptive Control of Thought — Rational") is the canonical hybrid architecture — symbolic and subsymbolic at once.

Take-away: modularity and PSSH-style processing can coexist with neural-net-style subsymbolic selection. Cognitive architecture is not all-or-nothing.

6.5 Worked example 1 — Ising network model of depression (Cramer et al. 2016)

Traditional latent-variable view: gallstones cause nausea, abdominal pain, heartburn — a single hidden cause produces all symptoms.

The network view of psychopathology: symptoms are nodes (active = 1, inactive = 0) coupled by weights W_ij. Activation propagates via a logistic function. Stress = extra input to all nodes.

• Depression evolves as a self-sustaining network of interacting symptoms.
• Insight into cognition without representations or computations.
• Same approach extends to bipolar disorder, generalized anxiety, attitude models.

6.6 Worked example 2 — Decision Field Theory (Busemeyer et al. 2019)

Choosing among multiple options (e.g., three phones differing in price, OS, battery, speed). Preference state P evolves over time:

The dynamics of a connectionist accumulator predict preferences, response times, and choice proportions as emergent properties of system evolution, not computations on symbols.

Take-away framing: "The behavior of the system as a whole is of interest — less focus on the computations on underlying representations, or even on architecture."

Bayesianism in Cognitive Science

Three ideas: (1) belief comes in degrees, (2) those degrees obey probability calculus, (3) learning = updating probabilities via Bayes' rule. The lecture's punchline: Bayesianism is the normative ideal — but humans systematically fail to reason like Bayesians.

7.1 The probability calculus rules

• Probabilities ∈ [0, 1].
• Impossible sentences = 0; necessary truths (2+2=4) = 1.
• If P and Q are logically equivalent, p(P) = p(Q).
• Negation: p(¬S) = 1 − p(S).
• Disjunction (mutually exclusive): p(R ∨ S) = p(R) + p(S).
• Conjunction (independent): p(R ∧ S) = p(R) × p(S).
• Conditional: p(A | B) = p(A ∧ B) / p(B).

7.2 Bayes' rule

The denominator is computed by marginalization: p(E) = p(E|H)·p(H) + p(E|¬H)·p(¬H)

7.3 Why the laws are objectively correct — Dutch books

A Dutch book is a set of bets that (1) the subject considers fair given their personal probabilities, but that (2) guarantee they lose money no matter what. Anyone whose beliefs violate probability calculus can be Dutch-booked. Therefore: rational degrees of belief must obey probability calculus.

Sam example: Sam believes "2+2=4" with probability 90%. Offer: he gets $0.90; he pays $1 if the bucket has 4 marbles. He thinks it's fair (EV = 0). But 2+2 IS 4 — so he always loses $0.10.

7.4 Worked example — ESP / clairvoyance

"Extraordinary claims require extraordinary evidence."— Carl Sagan (1978)

Clairvoyant correctly predicts 100 coin tosses. Should you believe in ESP?

• P(predict | ESP) = 0.9
• P(ESP) = 10⁻¹² (very skeptical prior)
• P(predict | ¬ESP) = 2⁻¹⁰⁰ ≈ 7 × 10⁻³¹

Posterior P(ESP | predict) ≈ 1 − 10⁻¹⁸. Almost certain — until you add the "trick" hypothesis with P(trick) = 10⁻⁶:

It's a million times more likely you were tricked than that ESP is real. The moral: tiny priors over fraud can outweigh enormous likelihoods.

7.5 Worked example — COVID-19 base-rate problem

Prevalence 1/1000. False positive rate 5%. You test positive. P(disease | positive) = ?

Reason over 1000 people:

• 1 person actually has it (positive).
• Of 999 healthy people, 5% test positive falsely ≈ 50.
• 51 positives total; only 1 is actually sick → ~2%, not 95%.

Why we still trust tests: in real life the sample is not random — there's a reason you were tested, so the relevant base rate is much higher.

7.6 The transposed-conditional fallacy

p(A | B) ≠ p(B | A).

• A = "is a white American man," B = "is a US senator"
• p(A | B) ≈ 0.9 (most senators are white American men)
• p(B | A) ≈ 0.00000009 (almost no white American men are senators)

📘 Textbook adds — Perception as Bayesian inference (Helmholtz → Hohwy)

The textbook frames Bayesianism in cognition as the modern formalisation of Hermann von Helmholtz's 19th-century proposal that perception is unconscious inference. The proximal sensory input radically underdetermines the distal world, so the brain must infer what is out there using stored knowledge about how the world tends to be.

• Hypothesis (H) = candidate layout of the distal environment.
• Evidence (E) = retinal stimulation.
• Likelihood p(E | H) = how probable this image is given that layout.
• Prior p(H) = how probable that layout is in general.
• Gestalt principles (continuity, proximity, good form, common fate) function as Bayesian priors over scene structure.

Case study: Binocular rivalry (Hohwy, Roepstorff & Friston, 2008)

Present a red iron to the left eye and a green violin to the right eye. Perception alternates between the two — never a stable composite. Why?

• H1 = red iron, H2 = green violin, H3 = composite "red-green iron-violin".
• Likelihoods of the conflicting retinal input are roughly equal across H1–H3.
• But p(H1) ≈ p(H2) ≫ p(H3) — the prior on composite objects is tiny.
• Posteriors: H1 and H2 tied, H3 ruled out. The visual system flips between the two equally-supported hypotheses rather than averaging them.

Binocular rivalry is a rational Bayesian response, not a glitch — a key example for predictive coding theories of perception.

7.7 Bayesian search theory (MH370)

For each grid cell i: p_i (probability object is there), a_i (probability of finding it if there), c_i (cost of searching).

After each miss, redistribute the posterior over remaining cells and recompute.

7.8 Where humans fail — heuristics & biases

7.9 The Linda problem — conjunction fallacy (Tversky & Kahneman)

Linda is 31, single, outspoken, very bright. Majored in philosophy; concerned with discrimination/social justice; antinuclear protests. Rank the probability:

• F. Linda is a bank teller
• H. Linda is a bank teller and active in the feminist movement

Most people rank H > F. But "feminist bank tellers" are a strict subset of "bank tellers." Specifying more detail can only LOWER probability, never raise it.

7.10 Bayesian view of psychopathology

Schizophrenic delusions can involve affirming the consequent: "Jesus had stigmata; I have stigmata; therefore I am Jesus." Rokeach (1964) "The Three Christs of Ypsilanti" — three paranoid schizophrenic men each believing he was Jesus, housed together for two years; their beliefs barely shifted, showing the difficulty of revising delusional priors.

Language Learning

Three paradigms applied to language: symbolic (Fodor's Language of Thought, Chomsky's innatism), connectionist (neural nets that reproduce children's overgeneralization), and Bayesian (statistical learning of word boundaries and anaphora). The lecture closes with the LLM revolution and what it implies about innateness.

8.1 What is language understanding?

• Semantics — meaning of words.
• Syntax — structure; surface vs deep structure.
• "Colorless green ideas sleep furiously" → syntactically well-formed, semantically anomalous.
• "John has hit the ball" / "The ball has been hit by John" → two surface structures, one deep structure.

Strong vs weak mastery: are linguistic rules explicitly represented in the head (strong sense, Fodor/Chomsky) or merely obeyed in behaviour (weak sense, connectionist/Bayesian)?

Key permission slip: "Rule-governed phenomena need not come from rule-governed information-processing structures." — this opens the door to connectionism and Bayesianism for language.

8.2 Fodor's Language of Thought (Mentalese)

Learning a language requires being able to evaluate truth conditions: "'The cat is on the mat' iff there's a cat and there's a mat and the cat is on the mat." Circularity problem: you can't learn this in English without already knowing what "cat" and "mat" are.

Solution: an innate symbolic medium — Mentalese. Slogan: "You cannot use the language you're learning to learn."

8.3 Nicaraguan Sign Language

In the 1970s, a school for deaf children in Nicaragua tried to teach Spanish by finger-spelling. Instead, the children spontaneously generated their own sign language. Documented by linguist Judy Kegl. Later generations added structural features like spatial modulation — evidence for innate language abilities.

8.4 Three paradigms for past-tense learning

The English past tense is a microcosm. Children show two features: (1) follow the "-ed" rule ("walked"); (2) handle exceptions ("gave"). Crucially, they make overgeneralization errors ("goed") that come and go in a gradual learning curve.

📘 Textbook adds — Rumelhart & McClelland (1986): the original past-tense network

Before Plunkett & Marchman there was the Rumelhart & McClelland (1986) PDP model — the founding connectionist past-tense network, published in the two-volume Parallel Distributed Processing.

The lineage matters because Pinker & Prince's critique of R&M is what motivated the dual-route symbolic model. The later Plunkett & Marchman result rebuts that critique on the connectionist side: overregularization emerges from co-presence of regulars and irregulars, not from training-set manipulation.

8.5 Bayesian language learning

(a) Word segmentation via transitional probabilities

For the sound string /k/ /ae/ /t/ /m/ /i/ /aʊ/ /z/ ("cat meows"):

• p(/ae/ | /k/) — high (within-word transition)
• p(/t/ | /ae/) — high
• p(/m/ | /t/) — low — this dip signals a word boundary

Same logic scales up to word-level transitional probabilities for sentence boundaries.

(b) Pronominal anaphora (Lidz et al.)

"I'll play with this red ball, and you can play with that one." Does "one" refer to H1 = a ball or H2 = a red ball?

• P(H | S) ∝ P(S | H) · P(H)
• Children learn P(S | H) from experience.
• Since P(S | H2) > P(S | H1), the most likely intended referent is "the red ball."

Reference: "What children know about syntax but could not have learnt."

8.6 LLMs — how they work

• Autoregression: feed each prediction back as input to predict the next.
• Transformer paper: "Attention is all you need" (Vaswani et al., 2017).
• Attention heads recode each token as a learned weighted combination of all tokens. Stacked hundreds of times, purely feedforward.
• The final encoding of the last token IS the prediction of the next token. Deterministic; randomness added post-hoc.

Embedding arithmetic — words as vectors

Two-stage training

1. Pre-training: mask the next word in billions of internet texts; backprop until predictions improve. (Tends to complete rather than reply.)
2. RLHF (Reinforcement Learning from Human Feedback): humans rate outputs; network updated toward higher-rated predictions. Makes models conversational.

LLM knowledge ≈ long-term semantic memory (fuzzy, can hallucinate); LLM prompts ≈ working memory (relevant info inserted reduces hallucinations).

8.7 Big debates — does this refute Chomsky?

Other critiques:

• Grounding problem — LLMs don't know what a banana tastes like.
• Mitchell & Krakauer (2023) — humans learn concepts; they abstract, reason compositionally and counterfactually, intervene on the world, and explain.
• Guest & Martin (2023) — multiple realizability: same outputs do not imply same mechanism.
• Training data asymmetry — GPT-3 saw ~4 × 10¹¹ words; a 5-year-old does causal reasoning on 4–5 orders of magnitude less input. Possible explanations: nativism, multi-modal grounding, active/social learning, or "comparing apples and pears."
• Binz & Schulz (2023, PNAS) — gave GPT-3 the Wason selection task. It got the canonical version right but failed ~50% of trials, with human-like error patterns.

Consciousness

Consciousness has two distinct dimensions (wakefulness and awareness), splits into easy vs. hard problems (Chalmers), and admits multiple competing theories. The empirical methods of cognitive science can address the easy problems; whether they can address the hard problem is a matter of fierce debate.

9.1 Two dimensions of consciousness (Laureys 2005)

In a study of 54 UWS/MCS patients, 5 could modulate brain activity via fMRI on command — they were conscious despite the clinical diagnosis.

9.2 The knowledge argument — Mary's Room (Jackson 1986)

"Mary is confined to a black-and-white room… she knows all the physical facts about us and our environment… It seems, however, that Mary does not know all that there is to know. For when she is let out of the black-and-white room or given a color television, she will learn what it is to see something red…"— Frank Jackson, 1986

Formal argument:

1. Inside the room, Mary has complete knowledge of how the brain processes colour.
2. So she knows everything about the information-processing of red.
3. When she leaves the room, she acquires new knowledge — what red is like.
4. Therefore, some aspects of conscious experience cannot be understood in terms of information processing.

9.3 Non-conscious processing — priming & dissociations

Strategy: contrast processing that works without awareness with processing that requires it.

• Face/Tool priming: categorization is faster when the prime is congruent (face primes face).
• Word priming: "DOG" is recognized faster after "CAT" than after "CAR" — the priming is semantic, not visual.
• Non-conscious priming is short-lived across SOA — consciousness allows information to be retained over time.
• Double dissociation — in disorder 1: A intact, B impaired; in disorder 2: B intact, A impaired. Strong evidence for separable mechanisms.

9.4 What is consciousness for?

Patients in blindsight or neglect never voluntarily act on stimuli in their affected field. The lecture's claim:

Consciousness permits identification of targets and planning of deliberate, voluntary action.

Milner & Goodale: two visual streams revisited

Fang & He (2005, Nature Neuroscience): interocular suppression renders a stimulus invisible. Result — robust dorsal activity even when the stimulus is invisible, but ventral activity tracks conscious perception. Conscious awareness is restricted to the ventral pathway.

9.5 Block's distinction

9.6 Chalmers — easy vs hard problems

9.7 Three theories of consciousness

📘 Textbook adds — Higher-Order Theories (Rosenthal, Armstrong, Lycan, Carruthers)

The textbook treats HOT as one of the major theories alongside GWT, IIT, and recurrent processing — the lecture only names it. Core claim: a mental state is conscious iff it is the object of a suitable higher-order mental state. The very same first-order state can be conscious at one moment and nonconscious at another, depending on whether something higher is "watching" it.

Standard objection to both: when I consciously smell a rose, I am aware of the rose, not of a mental state representing the rose — HOT/HOP seem to misdescribe the phenomenology. They also struggle to explain why consciousness has any distinctive functional role if first-order states behave identically with or without an accompanying HOT.

9.8 Dennett's deflationary view — the vitalism analogy

"The so-called hard problem of consciousness will disappear once we have a good enough understanding of the various phenomena lumped together under the label 'access consciousness'."— Daniel Dennett (1942–2024)

Analogy: in the 19th century, biology and chemistry seemed incapable in principle of explaining what separates living from non-living matter — there must be an élan vital ("vital force"). As biology matured, vitalism evaporated. Same fate (Dennett predicts) awaits the hard problem.

9.9 Cognitive scientists vs Mysterians

• Cognitive scientists: consciousness is a thriving research programme; the easy problems are tractable, and progress on them will eventually dissolve the hard problem.
• Mysterians: consciousness is, in principle, beyond cognitive-scientific tools.

Cross-cutting Themes

The lectures keep circling the same fault lines from different angles. Spotting them is half of cognitive science.

Theme 1 — Three paradigms (symbolic / connectionist / Bayesian-dynamical)

Theme 2 — Localization vs holism / encapsulation vs integration

• Phrenology (wrong in particulars, right in spirit) → Broca → Brodmann → fMRI subtraction → modular vision streams.
• Fodor: peripheral modules + central isotropic system. Massive modularity: all modules, no central system.
• Dorsal/ventral streams recur in lectures 3, 4, and 9 — different lesions, illusions, and consciousness studies all converge on the same anatomical dissociation.

Theme 3 — Conscious vs non-conscious processing

• Behaviourism would deny "the unconscious." But Corteen & Wood, blindsight, neglect, priming, and Owen et al.'s vegetative-state patients all show information processing without (or apart from) awareness.
• The lecture's unifying answer: consciousness is for deliberate, voluntary action and durable explicit information maintenance.

Theme 4 — Normative vs descriptive

• Bayesianism = how rational agents should reason.
• Tversky & Kahneman = how humans actually reason (badly, with predictable biases).
• Same tension in language learning: do children obey UG (norms) or are they statistical learners (descriptive)?

Theme 5 — The four big "dissolution" moves

1. Materialism dissolves dualism — Hobbes vs Descartes (lec 3).
2. S–S dissolves S–R — Rescorla vs Watson (lec 1).
3. Connectionism dissolves the need for explicit rules — Plunkett & Marchman vs dual-route (lec 5, 8).
4. Dennett dissolves the hard problem — vitalism analogy (lec 9).

Timeline of Key Figures & Events

Master Glossary

High-yield terms across all lectures. Skim the night before; nail them all.

Conditioning & learning

Brain & methods

Vision & categorization

Connectionism & AI

Bayes

Consciousness & mind

40-question Cross-lecture MCQ Practice

Mixed, harder, exam-shaped. Click "Show answer" only after committing. The choices include the most common distractors a professor will use.

— End of study guide —
Good luck on the exam. Trust the priors. Update on evidence.

The Emotions: From Cognitive Science to Affective Science

Emotions were largely ignored in early cognitive science. Affective science now studies them with a rich, multidisciplinary toolkit — from genetics and lesion studies to neuroimaging — using fear as its central case study.

16.1 Early Theories

Herbert Simon — Emotions as Interrupt Mechanisms

Simon (1967) argued that any sufficiently complex serial information-processing system (like the mind) must contain interrupt mechanisms — processes that can suspend an ongoing goal and substitute a new one when circumstances demand it. His proposal: emotions are those interrupt mechanisms in the CNS.

Three key properties of emotions on Simon's account:

1. They interrupt ongoing goals and substitute new goals/behaviours.
2. They arouse the autonomic nervous system in predictable physiological ways.
3. They generate feelings of emotion.

Simon's paper is influential but thin on detail — no specific emotion is actually analysed. It raises key open questions: How should emotions be classified? Are some basic? What is the role of arousal, physiology, and feeling? What are the neural bases?

Paul Ekman — Basic Emotion Theory

Ekman asked whether facial expressions of emotion are cross-cultural universals or products of social learning. To avoid the confound of media exposure, he studied the Fore linguistic-cultural group in New Guinea — a preliterate, visually isolated culture.

Method: participants were shown 2–3 photos of facial expressions while a story indicating an emotion was read aloud. They had to point to the matching face. Six target emotions: happiness, sadness, anger, surprise, disgust, fear.

Result: both adults and children in New Guinea matched emotions to faces at rates significantly above chance, supporting universal cross-cultural recognition. Ekman also showed that literate cultures could recognise New Guinean facial expressions.

Example — fear: eyebrows raised and horizontal, upper eyelid lifted, more sclera exposed (gathers information about threat); heart rate and skin conductance elevated; peripheral blood flow redirected to large skeletal muscles (preparing to flee).

Criticisms: Meta-analyses cast doubt on strict links between emotions and specific physiological/neural signatures. Cultural anthropologists question whether emotions are truly independent of social context.

16.2 Affective Space and the Affective Scientist's Toolkit

Affective Space: Beyond Discrete Categories

Affective phenomena vary in duration and function — emotions, moods, instincts, drives, and affective traits are all distinct (though their boundaries are fuzzy). Many affective scientists now model emotions as points in a multidimensional space rather than discrete categories.

The simplest model uses two dimensions:

• Valence — pleasure ↔ displeasure (attractiveness vs. aversiveness of a situation)
• Arousal — degree of physiological/psychological engagement

The circumplex model (Russell, 1980) plots emotions on a circle defined by valence and arousal. Includes both emotions and moods (e.g., sadness and depression are adjacent).

Adolphs & Anderson (2018) propose a richer 7-dimensional framework: scalability, valence, persistence, generalisation, global coordination, automaticity, social coordination. Designed to apply to non-human animals without relying on verbal self-reports.

Appraisal theories (Lazarus and others) add a cognitive dimension: emotions involve evaluating the environment relative to the subject's goals — what Lazarus calls core relational themes. On this view, anger and fear differ because they involve different appraisals of the same situation.

The Affective Scientist's Toolkit

Different tools study different components of an emotional episode (trigger → perception → neural/somatic response → behavioural response, with optional: cognitive appraisal, feelings, verbal report):

• fMRI, PET, electrophysiology — neural responses
• FACS (Facial Action Coding System) — behavioural/expressive responses
• Physiological measures — heart rate, skin conductance, finger temperature
• Lesion studies — causal role of specific brain regions
• Genetic tools — knockout experiments, optogenetics, pharmacogenetics

Genetic Tools

Knockout experiments (mice, 1990s+): replace a functional gene with a nonfunctional copy in stem cells, then study behavioural change. E.g., knocking out the serotonin receptor 5-HT(1A) increases anxiety-like behaviour in mice.

Optogenetics: engineer specific neurons to express a light-sensitive ion channel (opsin). Neurons can then be switched on/off with light (millisecond resolution). Allows targeted intervention in specific neural populations.

Pharmacogenetics: engineer neurons to express receptors for specific synthetic drugs (not normal neurotransmitters). Can activate or inhibit targeted populations. Slower than optogenetics (minutes to hours).

GECIs / GEVIs: genetically engineered indicators for calcium or voltage — allow optical measurement of neural activity as a complement to electrophysiology.

Lesion Studies

Classic case: Phineas Gage (1848) — iron rod through his head destroyed ventromedial prefrontal cortex. Physical and perceptual abilities preserved; emotional regulation and social behaviour drastically changed. First evidence linking prefrontal cortex to emotional function.

Lesions in humans provide information about dissociations; animal lesions allow greater anatomical precision and pre/post comparisons. Types: permanent (aspiration, excision, neurotoxins) or reversible (pharmacological, cryogenic cooling, TMS).

16.3 Fear: A Multilevel and Multidisciplinary Case Study

Fear Conditioning in Rodents

Fear conditioning pairs a neutral conditioned stimulus (CS — e.g., a tone) with an aversive unconditioned stimulus (US — e.g., foot shock). After training, CS alone elicits fear responses (freezing, autonomic arousal). This provides a controllable, reproducible way to study fear.

Multiple studies converge on the amygdala (a subcortical limbic structure) as central to fear:

• Electrical stimulation of the amygdala produces fear responses (increased respiration, heart rate, blood pressure, freezing) → amygdala is sufficient.
• Lesion studies: amygdala lesions abolish conditioned freezing and reduce unconditioned defensive behaviour (e.g., lesioned rats approach sedated cats) → amygdala is necessary.
• The basolateral nucleus (BLA) contains distinct neuron populations responding to positive vs. aversive stimuli, projecting to different brain regions.

Fear and Amygdala Damage in Humans — Patient S.M.

S.M. has bilateral amygdala destruction from Urbach–Wiethe disease (UWD), a rare genetic condition. Her basic cognition (intelligence, memory, language, perception) is intact. But she shows profound impairment in experiencing, expressing, and recognising fear.

Feinstein et al. (2011): exposed S.M. to live snakes and spiders, a haunted house, and scary films. She showed no fear responses — no avoidance, no subjective fear. She experienced other emotions normally, confirming fear-specific amygdala involvement.

Neuroimaging of Fear in Humans

LaBar et al. (1998): fMRI showed amygdala activation during fear conditioning and extinction in humans; greatest involvement in early stages of conditioning.

The human fear network identified by neuroimaging:

• Amygdala — central fear processing
• Hippocampus — memory/contextual information
• Insula — broader emotion processing (especially disgust)
• Anterior cingulate cortex (ACC) — bridges limbic and prefrontal systems
• Ventromedial prefrontal cortex (vmPFC) — integrative hub; modulates/controls fear responses (especially in extinction)

Clinical relevance: PTSD involves overgeneralised fear, slow extinction, amygdala/ACC hyperactivity, and low vmPFC activity.

Mobbs et al. (2010): scanned subjects while a live tarantula was placed at varying distances from their foot. Closer threat → increased amygdala, insula, ACC, BNST activation (active coping: flee). More distant threat → increased orbitomedial PFC activity (passive coping: freeze). Approach vs. retreat also differentially activated amygdala and BNST.

Summary

Affective science grew from cognitive science's neglect of emotion. Simon proposed a computational account (emotions as interrupt mechanisms); Ekman proposed discrete, universal basic emotions. Contemporary affective science has moved toward multidimensional models of affective space and uses a broad toolkit — neuroimaging, lesion studies, fear conditioning, and genetic tools. Fear is the best-studied emotion, with the amygdala (and its subregions BLA/CMA) playing a necessary and central role, embedded in a wider network including the hippocampus, insula, ACC, and vmPFC.