Week 4: Domain-General Learning in Infancy

What counts as learning?

Learning = a change in behavior or knowledge that LASTS and is CAUSED BY EXPERIENCE.

• Not learning:

  • Maturation: lasting, but caused by genes, not experience (e.g., puberty)

  • Transient changes: caused by experience, but temporary (e.g., most effects of drugs)

• TYPES:

  • asociative

  • statistical

  • error-driven

Associative learning

• Most basic form of learning: things that happen together = linked, one comes to trigger the other.

• Found in infants, adults, animals, even simple organisms.

  • Two types: classical conditioning and operant (reinforcement) learning.

Hebb's rule

Brain basis of associative learning.

• "Neurons that fire together, wire together."

• If neuron A repeatedly helps fire neuron B, A's connection to B gets stronger.

Classical conditioning

• WHAT: A neutral stimulus is paired with something that causes an INVOLUNTARY response, until the neutral stimulus alone causes the response.

• Example: Pavlov's dogs. A tone before food → eventually the tone alone causes salivation.

Goodman et al. 2018

Eye-blink conditioning

Question: Can infants learn by classical conditioning?

Method: A tone plays, then a puff of air hits the infant's eye.

Result: Eventually the tone alone makes the infant blink.

Claim: Infants learn by classical conditioning.

• Development: Newborns can do it too, but need a longer gap between tone and puff (~1.5 s vs. ~0.6 s), so they process more slowly.

Operant (reinforcement) learning

A VOLUNTARY behavior gets linked to its consequence.

• Reinforcement → behavior happens MORE

• Punishment → behavior happens LESS

• Difference from classical conditioning: classical = involuntary response (blink, salivate); operant = voluntary behavior (kicking, sitting).

Conjugate reinforcement

A type of reinforcement where the reward in proportion to the behavior. The more or harder the infant acts, the bigger the reward.

• Example: mobile kicking. The harder the baby kicks, the more the mobile moves.

Rovee and Rovee 1969; Rovee-Collier

Conjugate reinforcement with mobile kicking

Question: Can infants learn that their behavior causes an outcome, and remember it?

Method: A ribbon ties the baby's foot to a mobile. 4 phases: 1. Baseline: ribbon not attached → measure normal kicking 2. Acquisition: ribbon attached → kicking moves the mobile 3. Extinction: ribbon cut → kicking does nothing 4. Reacquisition: ribbon reattached

Result: Kicking goes up in acquisition, drops in extinction, and comes back FASTER in reacquisition.

Claim: Infants learn the contingency, and the faster relearning shows memory. Older = remembered contingency longer

Punishment (negative vs. positive)

• Negative punishment = REMOVE something good

  • Most studied in infants, because it's ethical

  • Ie: high-amplitude sucking - if the newborn sucks at the wrong rate, the mother's voice stops

• Positive punishment = ADD something bad

Papousek 1967

Question: Do infants learn from punishment?

Method: A beep signals: turn one way → milk; turn the other way → bitter liquid. After infants learn this, the sides are switched.

Result: Infants learned to turn toward the milk. After the switch, they struggled to stop and kept going to the now-bitter side.

Claim: Young infants struggle to stop a behavior that used to be rewarded. Punishment doesn't fully control them.

Real-world takeaway: children learn poorly from punishment. It says what is bad but leaves them wondering what is good. Better to reinforce the good behavior; if punishment is needed, use negative punishment (time-out). Spanking is ineffective and likely harmful.

Credit assignment problem

When a child is punished, they can't tell exactly what caused it.

• Example: punished for leaving the table. Was the table bad? The toys? The parent? One reason punishment is a poor teaching tool.

Statistical learning

Definition: picking up patterns (regularities) from the environment.

• Example: you can read text with no spaces because some letter pairs are common inside words and rare between words.

Saffran, Aslin and Newport 1996

Transitional probabilities

Question: Speech has no pauses between words. Can babies figure out where words start and end just from which syllables go together?

Method:

• Paradigm: familiarization + head-turn preference procedure

• Familiarization: 8-month-olds heard 2 min of nonstop made-up speech ("bidakupadotigolabu..."), secretly built from 4 made-up words. Inside a word, syllables always go together ("bi" → "da"); across words, rarely.

• Test (head-turn preference): a light flashes on one side of the room. When the baby turns to look at it, a sound plays from a speaker on that side, and keeps playing until the baby looks away. Looking time = listening time. Babies heard:

  • "Words" (bidaku): syllables that always went together

  • "Part-words" (kupado): end of one word + start of another

Result: Babies listened LONGER to part-words. What longer listening means: novelty preference. Part-words sounded less familiar → babies had learned the real words.

Claim: Babies track transition probabilities (how often one syllable follows another) and use them to find words.

Bulf et al. 2011

Question: Can newborns learn patterns in what they SEE, not just hear?

Method:

• Newborns (first weeks of life) watched shapes appear one at a time. - The shapes came in fixed pairs: square was always followed by X, circle was always followed by triangle (TP = 1, meaning 100% predictable).

• Test: two kinds of sequences:

  • Structured: the same pairs as before

  • Random: same shapes, but pairs broken up

Result: Newborns looked LONGER at the random sequence.

• What longer looking means: babies look longer at what's new or unexpected. The random sequence broke the pairs they'd learned → it looked new → so they had learned the pairs.

Claim: Visual statistical learning exists from the first weeks of life.

Critique (Addyman and Mareschal 2013): random sequences have more variety (less repetition). Maybe babies just like variety and didn't learn the pairs at all.

• Theme: how richly should we interpret infant behavior?

Flo et al. 2022

Question: Do newborns track syllable statistics automatically, even asleep?

Method: EEG on sleeping newborns while they heard a stream of made-up three-syllable words.

• Paradigm: EEG neural entrainment

Result: Their brain activity followed the rhythm of the words (neural entrainment).

Claim: Newborns group syllables into words using statistics, even during sleep.

Ellis et al. 2021

Hippocampus

Background: In adults, the hippocampus is critical for statistical learning. The infant hippocampus is still immature.

Question: Does the infant hippocampus still support statistical learning?

Method: Infant fMRI during structured vs. random sequences.

Result: No difference early in learning; a clear difference later in learning. Claim: The immature infant hippocampus responds to statistical patterns like an adult's.

Statistical learning applied to language

• Word segmentation: finding where words begin and end

• Word-label learning: the object in the center of a baby's view gets paired with its label over many experiences

• Reading: adults who are better statistical learners read faster

Bonus theme: infants match or beat adults on some statistical learning tasks → development ≠ improvement

Error-driven learning

What: improves performance by measuring and minimizing the difference between a predicted outcome and the actual ground truth

• Trial and error: act → notice the error → use its direction and size to adjust → try again.

• Example: force-field reaching (Krakauer 2009). A hidden force pushes your hand off course. You miss at first, then learn to aim off-center and hit, possibly without realizing it.

  • Why it's efficient: like video compression. Store only the differences (errors), not everything.

    • Related idea: predictive coding. The brain encodes only prediction errors.

Kidd et al. 2012

Goldilocks effect

Question: How much surprise do infants want?

Method: Objects popped out of boxes in sequences ranging from very predictable to very surprising.

• DV: when infants looked away.

Result: Infants looked away from events that were too predictable AND too surprising. They paid most attention to the middle (U-shaped curve).

Claim: Infants seek a "just right" amount of new information.

• Link to habituation: repeated things lose surprise (boring), so a new stimulus offers the right amount.

Stahl and Feigenson 2015

Question: Do infants explore in ways aimed at what surprised them?

Method: 11-month-olds saw a toy car do something impossible:

• Solidity violation: passed through a wall

• Support violation: rolled off a ledge and floated Then they got to play with the car.

Result:

• After the solidity violation → BANGED the car (testing if it's solid)

• After the support violation → DROPPED the car (testing gravity)

Claim: Infants test the specific expectation that was broken, like little scientists.

Predictive coding (Rao and Ballard 1999)

• the brain predicts then encodes only prediction errors, the difference between what happened and what it expected

• Efficient, like storing only the changes between video frames

Emberson et al. 2015

Question: Does the infant brain predict what's coming next?

Paradigm: fNIRS (a light-based cap that measures blood flow in the outer brain), with an omission trial.

Method:

• Background: a face activates the visual area (back of the brain); a sound activates the auditory area (side of the brain).

• Learning: 6-month-olds repeatedly heard a sound followed by a face, so the sound came to predict the face.

• Omission trial: the sound plays, but the face unexpectedly DOESN'T appear.

Result: On omission trials, the VISUAL area still activated, even though there was nothing to see.

Claim: The infant brain predicts upcoming events. It was preparing to see the face it expected.

Social learning

Learning from other people, for example by imitating what they do, to figure out the best behavior.

Perceptual learning (AKA: sensitization)

Definition: a long-lasting, experience-based improvement in perception

• The more you experience something, the better you perceive it.

• Examples: wine experts tell subtle tastes apart; birdwatchers identify birds from far away.

Key points:

• The change happens in the BRAIN, not the sense organ

• It doesn't depend on strategy or motor skill

Gomez et al. 2019

Pokemon

Question: Does childhood experience shape what the visual brain specializes in?

Paradigm: fMRI, comparing adult experts vs. novices (a natural experiment: childhood experience the researchers didn't control).

Method: Adults who played Pokemon a lot as kids (experts) vs adults who didn't (novices) viewed Pokemon pictures in a brain scanner.

Result: Experts had a distinct brain region that responded to Pokemon; novices didn't. Where it formed depended on how kids saw Pokemon: in central vision, on a small Game Boy screen. It sits near regions that respond to animals, not the face region.

Claim: Childhood experience shapes brain specialization (perceptual learning).

Petroff 2025

Oblique effect

Idea: The world has more horizontal and vertical lines than diagonal ones, so infants get more experience seeing them.

Finding (oblique effect): we detect changes in orientation better near vertical and horizontal than near diagonal.

Claim: Perception gets tuned to the statistics of what we see (perceptual learning).

Nakano and Kitazawa 2017

Question: When can infants remember an event they saw only once? (one-shot learning)

Paradigm: eye-tracking anticipatory looking, 24 hours later.

Method: 1.

• Day 1: infants (6–24 months) watched a video once: a person in a King Kong suit comes out of one of two identical doors and playfully attacks an actor.

• Day 2: they watched the same scene again. Before King Kong appears, do they look at the correct door?

Result: Only infants 18 months and older looked at the right door. 6, 12 and 15-month-olds did not.

Claim: One-shot learning (long-term memory of a single event) appears around 18 months.

Marcus et al. 1999

Question: Can infants do rule-based learning (learn abstract rules)?

Paradigm: familiarization, then the head-turn preference procedure

Method:

1. Familiarization: 7-month-olds heard 3-syllable sequences following an ABA pattern (le-di-le, wi-je-wi).

2. Test: NEW syllables they'd never heard, either following the same pattern (ba-po-ba) or a different one (ba-po-po).

Result: Infants listened longer to the different pattern (ba-po-po).

Original claim: Infants learned the abstract ABA rule, since the test syllables were new and only the pattern could tell them apart.

Critique: Later replications found the effect fragile or absent; The original may have been a fluke.

Theme: developmental science is an active research area.

• Marcus 1999 vs. the failed replications

Inductive inference

Definition: generalizing from a SAMPLE to a whole POPULATION (or the reverse). Going from a few examples to a broader conclusion.

• Part of "children as scientists": children build theories, test them, and revise them.

• Faster than statistical learning, which needs lots of input.

Ruff 1978

3D shapes

Question: Can infants figure out an object's full 3D shape from just a few views?

Paradigm: familiarization, then novelty preference (looking time).

Method:

1. Familiarization: infants saw a 3D shape shown from several different angles.

2. Test: the SAME shape from a NEW angle vs. a DIFFERENT shape.

Result: Infants looked longer at the different shape, so they recognized the original even from an angle they'd never seen.

Claim: Infants can infer new viewpoints of an object.

Limit: until about 5 months, they need several views during familiarization; one view isn't enough.

Wood and Wood 2020

Chicks

Question: What experience do newborn animals need to build a 3D object from different views?

Paradigm: controlled rearing (chicks raised in controlled environments from hatching).

Method: Chicks saw an object's different viewpoints either as smooth, continuous motion (like the object rotating) or as separate images (more like stop motion).

Result: Chicks recognized the object as one 3D thing only when they saw the views in continuous motion.

Claim: Continuous motion is a key cue for building 3D object representations.

Xu and Garcia 2008

Question: Can infants reason from a sample to a population, and from a population to a sample?

Paradigm: violation of expectation (looking time).

Method:

1. Sample → population: 8-month-olds watched an experimenter pull a handful of balls out of a covered box, without looking inside. The sample was mostly red (e.g., 4 red, 1 white) or mostly white. Then the box was opened to show its contents: mostly red or mostly white.

2. Population → sample (reversed): infants first saw what was in the box, then watched a sample being drawn out.

Result: Infants looked longer when the sample and box didn't match (e.g., a mostly-red sample from a mostly-white box), in both directions. What longer looking means: surprise. The mismatch was improbable.

Claim: Infants make probability-based inferences from sample to population and back.

Causal reasoning

What: A makes B happen, causation

Problem: Constant Conjunction

• we never see causation directly, only that A and B happen together

  • Just because they happen together doesn’t mean they are casual

Leslie and Keeble 1987

Launching

Question: Do 6-month-olds perceive cause and effect, or just motion?

Paradigm: habituation, then dishabituation at test (looking time).

• babies stare longer at things that are new/suprising

Method:

• Two groups, habituated to different films:

  • Direct launching: ball A hits ball B, and B moves off IMMEDIATELY (looks causal to adults)

  • Delayed launching: A hits B, a short PAUSE, then B moves (doesn't look causal)

• Test: each group sees its film played BACKWARD.

  • Logic: only the direct-launching reversal also swaps cause and effect (agents change and now B seems to launch A).

Result: Babies looked longer at the reversed DIRECT launching film.

Claim: 6-month-olds perceive causal structure, not just movement.

Kushnir and Gopnik 2007

Question: Do children assume causes must touch their effects, and can evidence change their minds?

Paradigm: "blicket detector" task. A machine lights up and plays music when certain blocks are placed on it; children choose how to make it go.

Method: Preschoolers saw blocks activate the machine. The evidence showed that activation worked better either when the block TOUCHED the machine (contact condition) or when it HOVERED above it (no-contact condition).

Result:

• Evidence overrides physical bias: Preschoolers readily use new statistical data (conditional probability) to override their deep-seated assumption that a cause must physically touch an effect to work.

• Prior expectations act as a filter: While children can change their minds based on data, they learn much faster and make more robust causal inferences when the evidence aligns with physical contact than when it contradicts it.

Claim: Preschoolers assume causes work by contact, but can learn otherwise from evidence. Action at a distance stays harder.

Bonawitz et al. 2010

Question: If toddlers can PREDICT that A causes B, will they ACT to make A happen?

Paradigm: anticipatory looking (prediction), then an intervention test (action).

Method:

1. Children watched a block move and touch another block; then a toy (a fan) started spinning. This repeated.

2. Prediction: did children look at the fan before it spun? (Shows they learned A → B.)

3. Intervention: children got the block. Would they move it to make the fan spin?

Result: -

• 2-year-olds PREDICTED the fan (looked ahead) but didn't move the block to make it go.

• 4-year-olds did both.

  • Toddlers succeeded only when a person (agent) moved the block, when the objects touched directly, or when the experimenter used causal language ("the block can make it go").

Claim: Predicting a causal relation and acting on it are different. Toddlers have a gap between them.

Gweon and Schulz 2011

Question: When something fails, can infants tell whether it's their fault or the object's?

Paradigm: behavioral choice.

• DV: does the baby hand the toy to a parent (asking for help) or reach for another toy (exploring)?

Method:

• Experiment 1: A 16-month-old sits next to a parent. The experimenter presses a button on a green toy, and it plays music. A red toy sits nearby on a cloth.

  • Same-toy condition: baby gets the SAME green toy. It doesn't work for them. It just worked for the adult, so the likely problem = the baby.

  • Different-toy condition: baby gets a YELLOW toy (same toy, different color). It doesn't work. Different toy, so the likely problem = the toy.

• Experiment 2: baby always got the same green toy, but what they saw beforehand changed: one adult always succeeded and another always failed (the person matters), or each adult succeeded once and failed once (the toy is unreliable).

Result:

• Evidence pointing to the person (same toy, or adult-dependent success) → handed the toy to the parent

• Evidence pointing to the toy (different toy, or unreliable toy) → reached for the red toy

Claim: 16-month-olds use evidence to figure out the cause of a failure, and act on it: ask for help if it's them, try something else if it's the toy.

Disjunctive syllogism

Strict rule of formal logic used to eliminate options

• Reasoning by elimination

1. A or B

2. Not A

3. Therefore B

• Example: either the cafe has a long line or it's closed. It isn't closed. So there's a long line.

Cesana-Arlotti et al. 2018

Question: Can babies who don't talk yet reason by elimination?

Paradigm: violation of expectation, with eye-tracking and pupil dilation.

Method:

1. Two different objects (e.g., a ball and snake) go behind a wall.

2. A cup scoops one of them out, but babies can't tell which, because only its top edge shows. (Either A or B is in the cup.)

3. The wall drops, showing object A still behind it. (So it's NOT A.)

4. Logic: the cup must hold B.

5. Then the cup's contents are revealed.

Result:

• Babies looked longer when the outcome contradicted the logic (e.g., the cup didn't hold B).

• Their pupils dilated more in scenes that required a deduction, as adults' do when reasoning.

Claim: 12-month-olds may use logical deduction (process of elimination) before they have language.

Counterfactual reasoning

Definition: imagining a different past and its outcome:

• "If A HADN'T happened, would B still have happened?"

• Infants: probably can't do it.

• It develops slowly.

• Debate: Harris 1996 (young children can) vs. Rafetseder 2013 (not until ~12–14 years)

Harris et al. 1996

Question: Can young children reason about what WOULD have happened?

Paradigm: verbal story task (a counterfactual "what if" question).

Method: Story with a doll: Carol walks into a house with a clean kitchen floor. Her shoes are muddy; she could take them off but doesn't, and leaves muddy footprints. Children are asked: "What if Carol had taken her shoes off? Would the floor be dirty?"

Result: 3–4-year-olds correctly said no, it would be clean.

Claim: Young children can mentally undo an event and imagine the different outcome.

Rafetseder et al. 2013

(opposes Harris 1996)

Question: Were Harris's kids really reasoning counterfactually, or just using a simple rule?

• The problem with Harris: kids could answer correctly with "basic conditional reasoning," a general rule like "no muddy shoes → clean floor," without actually imagining the specific story going differently.

Paradigm: Harris-style story questions, but designed so the simple rule gives the WRONG answer. (E.g., the floor also got dirty another way, so even without the muddy shoes it would still be dirty. True counterfactual reasoning says "still dirty"; the rule says "clean.")

Result: 9–11-year-olds did no better than 6-year-olds. Only 12–14-year-olds came close to adults.

Claim: True counterfactual reasoning develops slowly, not until ~12–14 years.

• Opposing Harris 1996

Theme: how richly to interpret children's answers.

Summary of infant reasoning

Inductive

• Can: reason from sample to population and back (Xu and Garcia)

• Limit: need more examples than adults to generalize

Causal

• Can: notice when cause and effect are reversed (Leslie and Keeble)

• Limit: assume causes need contact; evidence can override this, but action at a distance stays harder (Kushnir and Gopnik)

Deductive

• Can: possibly reason by elimination at 12 months (Cesana-Arlotti)

• Limit: how to interpret it is debated

Counterfactual

• Can: little or nothing

• Limit: develops slowly, ~12–14 years (Rafetseder)

Faces: domain-specific vs. domain-general

Question: Are babies born with a special system for faces?

Domain-specific view: born with a face template; brain regions specialized for faces; experience fine-tunes it.

• Evidence:

  • Goren 1975: newborns follow face-like patterns farther

  • Farroni 2005: newborns prefer upright face-like patterns, but only with dark eye and mouth areas

  • Farroni 2002: newborns prefer eyes looking straight at them

  • Fusiform face area (FFA): face-selective brain region, even in blind adults touching faces

Domain-general view: born with only general visual biases; experience builds face processing.

• Evidence:

  • Pascalis 1995: mom-face preference disappears with headscarves (hairline, not face)

  • Sai 2005: mom-face preference needs hearing her voice

  • Arcaro 2017: monkeys raised without seeing faces develop no face areas

Reconciliation: perceptual narrowing. Babies lose sensitivity to faces they don't see.

• Pascalis 2002: 6-month-olds tell monkey faces apart; 9-month-olds can't

• Scott and Monesson 2009: naming individual monkey faces keeps the ability

Statistical vs. associative learning

Statistical learning differs from associative learning in 3 ways:

Reward: associative usually needs a reward/punishment; statistical doesn't (happens in the background)

Regularity: associative needs near-perfect pairing (~90%); statistical works with weak patterns (~20%)

Awareness: associative can be conscious; statistical is largely implicit

Transition probability (TP)

How likely one item (syllable, shape) is to follow another.

High inside words (bi → da = 100% in Saffran 1996)

Low across word boundaries (~25%)

Dips in TP mark where words begin and end → used for word segmentation

Krakauer 2009

Force-field reaching (error-driven learning)

Method: Move a joystick to a target while a hidden force field pushes your hand off course.

Result: Miss on trial 1; by ~trial 100 you aim off-center and hit the target, possibly without realizing it.

Claim: We use the direction and size of errors to adjust, even without awareness.

Dark room problem (Sun and Firestone 2020)

If we ONLY minimized prediction error, we'd hide in a dark room and never learn.

So learners also SEEK surprise → Kidd 2012 (Goldilocks effect), Stahl and Feigenson 2015 (exploration)

Anticipatory looking

After learning a sequence (e.g., green then purple), infants look to where the next item will appear BEFORE it appears.

Shows prediction without needing other behavior.

Used in: Nakano and Kitazawa 2017 (one-shot memory), Bonawitz 2010 (causal prediction)

Addyman and Mareschal 2013

Critique of Bulf 2011 (opposing pair)

Argument: Random sequences have less repetition (less "local redundancy") than structured ones.

So newborns may look longer at random sequences just because they prefer VARIETY, not because they learned the pairs.

Theme: how richly should we interpret infant behavior?

Capacities infants probably DON'T have

Imprinting

One-shot learning (until ~18 months)

Rule-based learning

Counterfactual reasoning (until ~12–14 years)

AI example: backpropagation (main neural network algorithm) isn't practical for brains → every new piece of info would mean updating the whole network

Imprinting

Definition: becoming attached to whatever is seen right after birth, within a short critical period.

Example: ducklings/chicks follow the first thing they see (Regolin and Vallortigara)

Related: genetically triggered learning, e.g., songbirds learning their species' song (Louder et al. 2019)

Humans: decades of research find NO imprinting; the first faces a newborn sees aren't special.

One-shot learning

Definition: remembering something after a SINGLE exposure. Needed for episodic memory (e.g., remembering your 10th birthday).

Infants: can remember faces for 24 hours, but only after MANY exposures (Pascalis et al. 1998)

One-shot memory appears ~18 months (Nakano and Kitazawa 2017)

Rule-based learning

Definition: extracting abstract patterns that generalize beyond surface details (reasoning by analogy).

Evidence: Marcus 1999 claimed 7-month-olds learn ABA rules

But: failed replications (Rabagliati et al. 2019; Geambasu et al. 2023) → effect fragile or absent

Theme: developmental science is an active research area (in 2018 this lecture would have said infants CAN learn rules)

Children as scientists

Children build theories about how the world works, test them, and revise them based on evidence.

Types of reasoning:

Inductive (sample → population)

Causal (A makes B happen)

Deductive (certain conclusions from premises)

Counterfactual (what would have happened)

Deductive inference

Definition: reasoning with CERTAINTY from premises to a conclusion.

Example form: disjunctive syllogism (A or B; not A; therefore B)

Infants: possibly at 12 months (Cesana-Arlotti 2018), but interpretation is debated

Goren 1975

Newborns track faces (domain-specific evidence)

Question: Do newborns prefer faces from birth?

Method: Visual tracking. Newborns (minutes old) watched a paddle moved slowly across their view showing a simple drawn face, a scrambled face, or a blank.

Result: Newborns turned their eyes and head FARTHER to follow the face.

Claim: Newborns prefer faces from birth.

Farroni 2005

Face-like patterns (domain-specific evidence)

Question: What makes newborns prefer face-like patterns?

Method: Preferential looking. Three blobs arranged like a face (2 eyes on top, mouth below), upright vs. upside down. Versions with darker eye/mouth areas (like real faces) vs. light and dark reversed.

Result: Newborns preferred the upright pattern ONLY when eye and mouth areas were darker. Reversed contrast → preference disappeared.

Claim: The newborn face bias is tuned to real-face features (dark eyes and mouth), possibly for detecting eye contact.

Farroni 2002

Eye gaze (domain-specific evidence)

Method: Preferential looking. Newborns saw a face looking straight at them vs. a face with averted gaze.

Result: Newborns preferred direct gaze.

Claim: Early sensitivity to eye gaze.

Fusiform face area (FFA)

Face-selective brain region (domain-specific evidence)

Method: fMRI

Finding: Responds selectively to faces in adults and infants, and even in congenitally blind adults when they TOUCH faces.

Claim: The brain has regions specialized for faces.

Pascalis 1995

Mom's face or her hairline? (domain-general evidence)

Question: Do newborns recognize their mother's face?

Method: Preferential looking. Newborns (a few days old) saw mom's face next to a stranger's with similar hair and coloring. Then both women wore headscarves covering hair and hairline.

Result: Newborns looked longer at mom, but the preference DISAPPEARED with headscarves.

Claim: Newborns may recognize the hairline, not the face.

Sai 2005

Mom's face needs her voice (domain-general evidence)

Method: Preferential looking (mom vs. stranger). In one group, mothers didn't talk to their babies from birth until testing, so babies saw mom's face without hearing her voice.

Result: Babies who had seen AND heard mom preferred her face; babies who only saw her silently showed no preference.

Claim: Face preference depends on experience (hearing her voice).

Arcaro 2017

Face-deprived monkeys (domain-general evidence)

Question: Do face areas need face experience?

Method: Controlled rearing. Baby monkeys raised by caretakers wearing welding masks (~first 200 days), so they never saw a face. Then fMRI and gaze tracking.

Result: No face-specialized brain regions developed (other categories developed normally). They also didn't look preferentially at faces, unlike normally raised monkeys.

Claim: Face areas need face experience; even the tendency to look at faces isn't innate.

Theme: nature and nurture

Perceptual narrowing (faces)

Reconciles the domain-specific vs. domain-general face debate.

Infants lose sensitivity to faces they DON'T experience.

Evidence: Pascalis 2002 (monkey faces), other-race effect, Scott and Monesson 2009 (training prevents it)

Theme: development does not equate to improvement

Pascalis et al. 2002

Monkey faces

Question: Does face perception narrow with experience?

Method: Visual paired comparison. Babies see one face (human or monkey) for a while, then that face next to a new face of the same species. Looking longer at the new face = can tell them apart.

Result: 6-month-olds told apart BOTH human and monkey faces. 9-month-olds and adults only human faces.

Claim: Face perception narrows with experience.

Theme: development does not equate to improvement

Other-race effect

Infants come to discriminate faces of their own race better than faces of other races.

Claim: Narrowing follows what infants see.

Scott and Monesson 2009

Escaping face narrowing

Method: Training intervention. 6-month-olds were trained with individual names for monkey faces ("this one likes berries, this one likes bananas"). Tested at 9 months with visual paired comparison.

Result: Trained infants KEPT the ability to tell monkey faces apart.

Claim: The right experience (individual-level labels) prevents narrowi