Week 3: Domain-Specific Knowledge in Infancy

Nature vs. nurture

Nature: capacities and developmental events programmed by genes

Nurture: the environment and experiences we respond to and change from

Modern view: NOT one vs. the other. They work together; the question is which traits depend more on each.

Today the debate = domain-specific vs. domain-general

4 ways to test what comes from nature

1. Test very young infants: if they already know it, they probably weren't taught it

2. Compare with animals: shared with a common ancestor → genetically programmed; shared with an unrelated species → convergent evolution (same pressures, same solution)

3. Compare across cultures: universal trait → nature

4. Twin studies

Hershberger (controlled-rearing chicks)

Light from above

Question: Is the assumption that "light comes from above" learned or innate?

Method: Controlled rearing. Chicks raised where light came ONLY from below. Then shown two shaded circles: lit on top (looks like a bump) vs. lit on bottom (looks like a dent).

Result: Chicks still treated the top-lit circle as sticking out, despite their own experience.

Claim: "Light from above" is built in.

Theme: nature

Ekman

Emotions across cultures

Question: Are facial expressions of emotion learned from culture or universal?

Method: Cross-cultural comparison. Studied isolated tribes with no contact with Western culture.

Result: Their facial expressions matched Western ones.

Claim: Basic emotional expressions are universal → points to nature.

Baldwin effect

Behavior that helps survival can shape genetic evolution over generations.

Learning itself isn't inherited, but traits that SUPPORT the learning are.

Examples:

Better hand dexterity survived → supported more toolmaking

Language: vocalizing came first, then selection favored better vocal cords

Epigenetics

The environment changes which genes are EXPRESSED, without changing the genetic code itself.

Example: Weaver et al. 2004 (maternal grooming in rats)

Intergenerational transmission of trauma

Effects of trauma can carry to the next generation.

Genetic changes play only a small role; environmental factors (parenting, poverty) matter more.

Open research area.

Weaver et al. 2004

Epigenetics in rats

Question: Can a mother's care change her offspring's genes?

Method: Mother rats naturally differ in how much they lick and groom pups. Compared offspring of high- vs. low-grooming mothers: measured methylation (chemical tags) on a stress-hormone receptor gene in the hippocampus, and stress reactions. Cross-fostering (swapping pups between mothers) showed the effect came from the care received, not inherited genes.

Result: Pups of low-grooming mothers had more methylation on that gene → gene turned down → fewer stress-hormone receptors → stronger stress response.

Claim: Experience changes gene expression, so nature and nurture interact.

Theme: nature and nurture

Domain-specific vs. domain-general

Both agree innate capacities AND experience matter; they disagree on how much structure is built in.

Domain-specific: infants are born with fragile, innate knowledge in specific domains (objects, numbers, language). Slogan: the learning problem is too big without built-in structure. Evolution: specialized mechanisms for specific problems. → Week 3 (core knowledge)

Domain-general: infants are born with general learning tools (attention, memory, inference) that apply to everything. Slogan: learning is all you need (good learning rules + lots of data). Evolution: general mechanisms for varied environments. → Week 4 (learning mechanisms)

Domain-specific vs. general in AI

Does human-like AI need built-in structure (like core knowledge), or can it emerge from data + learning rules alone?

Labs (including the lecturer's) train AI on video/audio recorded from infants' daily lives to see what can be learned with only domain-general assumptions.

So far: models learn a lot without built-in structure, but may still be missing something. Open question.

Core knowledge theory (Elizabeth Spelke)

Infants are born with innate knowledge in a few domains: objects, numbers, places, agents (+ forms and social reasoning).

These are building blocks that later learning builds on.

Predictions:

Detectable as soon as they can be measured

Present in other animals and across the lifespan (from evolutionary past)

Have signature limits

Information is encapsulated

Signature limits

Core knowledge systems FAIL under simple variations evolution didn't prepare them for.

Examples:

A chicken freezes and stares at a V drawn in the dirt (line-detecting system tricked)

Humans tricked by sugary food, social media, reinforcement loops

Infant examples: exact number fails above 3 (Feigenson crackers); reorientation ignores landmarks (Hermer and Spelke)

Encapsulation

Some processing happens automatically, no matter what we know.

Example: pareidolia (seeing faces in objects)

Core knowledge of objects: 4 principles

Cohesion: objects move as connected, bounded wholes

Continuity: objects persist in space and time; they don't vanish or teleport

Contact: objects don't move by themselves; only when something contacts them

Support: unsupported objects fall; they can't float

Piaget's object permanence task

The old view

Question: When do infants understand that hidden objects still exist?

Method: Manual search (behavior-heavy). While the infant watches, a toy is covered with a cloth. Does the infant lift the cloth to get it?

Result: Infants don't search until ~9–10 months.

Claim: Infants can't represent hidden objects until ~10 months.

Critique: Searching takes motor coordination. Looking-time studies (only need visual attention) show infants represent hidden objects months earlier.

Themes: reducing demands on behavior; active research area

Kellman and Spelke 1983

Rod and box (amodal completion)

Question: Do infants perceive a partly hidden object as whole?

Method: Habituation, then looking time. Habituate to a rod moving back and forth behind a box (middle hidden). Then remove the box: show one complete rod or two broken pieces.

Result: Infants looked longer at the BROKEN rod (unexpected).

Claim: Infants fill in the hidden part (amodal completion) and treat the rod as one continuous object.

Baillargeon drawbridge (solidity)

Question: Do infants expect objects to be solid?

Method: Habituation + violation of expectation. Habituate to a rotating screen (drawbridge) that is stopped by a block. At test, it rotates right through the space where the block should be.

Result: Infants look longer at the impossible event.

Claim: Infants expect objects to be solid.

Related findings: infants look ahead to where a ball rolling behind a box should reappear (continuity); surprised when an object rolls off a table and floats (support)

Spelke 1992

Ball and shelf

Question: Do infants expect a falling object to be stopped by a solid surface?

Method: Habituation + VoE. Habituation: a ball drops behind a screen; screen lifts → ball on the floor. Test: a shelf is added above the floor. Consistent: ball on the shelf. Inconsistent: ball on the floor below the shelf (would have passed through it).

Result: Infants looked longer at the inconsistent event, even though it looked MORE like habituation.

Claim: Infants expect solidity and continuity.

Xu and Carey 1996

Duck and truck

Question: Can infants use WHAT an object is to work out how many objects there are?

Method: VoE. A duck comes out from one side of a screen and goes back; then a truck comes out the other side and goes back (never both at once). Screen removed → 2 objects (expected) or 1 (unexpected). Controls: baseline looking at 1 vs. 2 objects; condition where both objects were shown together at the start.

Result: 10-month-olds were NOT surprised by 1 object (accepted a duck "turning into" a truck). 12-month-olds were surprised. 10-month-olds DID succeed when they saw both objects at once (they can use location/movement, just not object kind).

Claim: Early object knowledge has limits: younger infants don't use an object's kind or features to track separate objects.

Feigenson and Carey 2003

Reaching into a box

Method: Infants watched small objects go into a box they could reach into but not see inside. Some objects were secretly removed. DV: how long they keep searching after pulling some out.

Result: 12–14-month-olds searched correctly with up to 3 objects (saw 3 go in, got 2 out → kept searching), but failed with 4.

Claim: Same limit of 3 as the crackers study (signature limit of exact number).

Wynn 1992

"1 + 1"

Infants are surprised when the number of objects behind a screen changes by "magic" (e.g., 1 + 1 = 1).

Signature limit: works for solid objects but breaks down for non-cohesive things like piles of sand.

Designing around critics

Critics look for confounds:

Baseline preferences (infants may just like one display more)

Salience (some events are simply more interesting)

Researchers answer with control conditions: test baseline looking with no surprising event; keep conditions identical except the surprising part.

This back-and-forth is how the field improves.

Object knowledge in chicks

Imprinting: chicks imprint on objects they see right after hatching → used to test what chicks treat as an object.

Regolin and Vallortigara 1995 (hidden triangle): chicks imprinted on a moving triangle partly hidden by a square. Afterward they preferred a COMPLETE triangle over its separated pieces → they "filled in" the hidden part.

Controlled-rearing chicks that never saw solid objects still assume objects are solid.

Takeaway: object knowledge reflects basic intuitive physics and appears across species.

Exact number (subitizing) vs. approximate number system (ANS)

Exact number (subitizing): instantly knowing how many items there are, without counting. Limit: ~3–4 items. Infants: yes, up to 3.

Approximate number system (ANS): estimating which group has more, without exact numbers. Ratio-limited (10 vs. 20 is as easy as 20 vs. 40). Infants: yes, ratio-sensitive like adults (~1:2 at 6 months).

Feigenson, Carey and Hauser 2002

Crackers

Question: What is the limit of infants' exact number tracking?

Method: Manual choice. 10–12-month-olds watch crackers dropped one at a time into two opaque buckets (e.g., 1 left, 2 right), never seeing all at once. Then they crawl to a bucket. Pairs: 1 vs. 2, 2 vs. 3, 3 vs. 4, 2 vs. 4, 1 vs. 4.

Result: Chose the bigger amount for 1 vs. 2 and 2 vs. 3. At chance for 3 vs. 4, 2 vs. 4, and even 1 vs. 4.

Claim: Exact tracking has a signature limit of 3 items; it breaks down beyond 3, even for a big difference like 1 vs. 4.

Agrillo 2021

Guppies

Question: Do fish have number representations?

Method: Spontaneous choice. Tank in 3 sections: 2 guppies on one side, 3 on the other, test guppy in the middle. Guppies prefer bigger groups → where does it spend time? Then larger ratios (e.g., 4 vs. 16).

Result: Only a slight preference for 3 over 2 (weak exact number). Clear preference for the bigger group at large ratios.

Claim: Even fish have an approximate number system → evolutionarily old.

Brannon 2006

Number vs. area

Finding: 6-month-olds notice changes in area but are MORE sensitive to changes in number.

By 6 months, infants detect a change in the number of dots on a screen.

The area problem

Studies try to control total area so number isn't confounded with "how much stuff."

But 2022 research suggests the brain may not compute true area; it may use simpler measures (perimeter, or height × width = bounding box).

So controlling area doesn't fully remove the confound → need humility in interpreting results.

Place cells and path integration

Place cells: neurons in the hippocampus that fire at specific locations (found in rats and other animals)

Core knowledge view: innate capacity to represent where we are and where we can go

Path integration: tracking your position from your own movements (direction and distance), without vision

Wittlinger, Wehner and Wolf 2006

Desert ants (path integration)

Question: How do desert ants find their way home without scent trails?

Method: Field experiment. After ants walked to a feeder, researchers glued on stilts (longer legs) or cut legs to stumps (shorter). Another version: ants picked up and moved after finding food.

Result: Ants on stilts walked PAST the nest; ants on stumps stopped SHORT. Moved ants searched where the nest would have been from their original position.

Claim: Ants use path integration, measuring distance by counting steps. Navigation by internal computation is evolutionarily very old.

Landau, Gleitman and Spelke 1981

Blind child navigating

Question: Do you need vision to build a mental map?

Method: Case study. A blind child was guided between some objects in a room, then asked to go directly between objects along routes she had never traveled.

Result: She found the new routes. Blindfolded sighted people can do similar tasks.

Claim: Spatial navigation doesn't depend on vision; it's an early, robust capacity.

Hermer and Spelke 1994

Reorientation

Question: What do children use to reorient after getting lost?

Method: Disorientation + search. In a rectangular room, children see a reward hidden in one corner. They're spun around, placed in the center, and allowed to search. DV: which corner they search.

Result: Searched equally at the correct corner and the geometrically identical opposite corner, even when one wall was a distinctive color.

Claim: Children reorient using room GEOMETRY. Signature limit: they ignore a useful landmark.

Agents

Like objects (cohesive, continuous) but can move on their own and have GOALS.

Adults over-attribute agency (moving triangles and circles seem to have intentions).

Agents can create order out of disorder; non-agents can only increase disorder.

Biological motion: agents have a typical way of moving. Newborns look longer at biological than non-biological motion.

Johnson, Slaughter and Carey 1998

What counts as an agent?

Question: What cues make infants treat something as an agent?

Method: Gaze following, 2 x 2 design (face or no face x contingent or not). A fuzzy blob sat in front of a 12-month-old. It either had a face (eyes) or not, and either responded to the baby (beeping/lighting up when the baby babbled or moved = contingent) or acted on its own schedule. Then the blob "turned" toward one of two toys. DV: did the baby look where the blob turned?

Result: 12-month-olds followed the blob's "gaze" when it had a face OR behaved contingently.

Claim: Infants use either a face or contingent behavior to decide something is an agent.

Gergely and Csibra 1995

Efficient paths

Question: Do infants expect agents to act efficiently?

Method: Habituation, then looking time. Habituation: an agent jumps over a barrier to reach a goal. Test: barrier removed; agent either jumps the same way (inefficient) or goes straight (efficient).

Result: 12-month-olds looked longer at the inefficient jump, even though it looked the same as habituation.

Claim: Infants expect agents to pursue goals efficiently.

Costly actions and Woodward 1998 (goal switch)

Costly actions: infants infer that an agent who works harder for a goal values it more.

Woodward 1998 (goal switch): 5-month-olds look longer when an agent reaches for a DIFFERENT goal than before → they encode the goal, not just the movement.

Skerry 2013

Sticky mittens

Question: Can infants who can't yet reach understand efficient reaching, if given reaching experience?

Method: Training intervention, then habituation + looking time. 3-month-olds wore Velcro "sticky mittens" so toys stuck when they swiped (reaching experience, nothing in the way). Then habituation: actor reaches OVER a barrier for a toy. Test: barrier removed; actor reaches straight (efficient, new motion) or makes the same arched reach (inefficient, same motion). Compared with untrained infants.

Result: Trained infants looked longer at INEFFICIENT reaches; untrained infants did not.

Claim: Reaching experience helps infants see an agent's goal. The efficiency expectation goes beyond the training (they never practiced around barriers), so it seems already there, as core knowledge predicts.

Themes: reducing demands on behavior; nature and nurture

Newman 2010

Agents create order

Question: Do infants know only agents can create order?

Method: VoE. A screen hides blocks while a ball (non-agent) or a self-moving character with eyes (agent) goes behind it. Screen drops: blocks went messy → neat (order) or neat → messy (disorder).

Result: 12-month-olds looked longer when the NON-AGENT created order. 7-month-olds showed no difference. With the agent, looking was about equal (agents can do either).

Claim: By 12 months, infants link agents to creating order.

Critique: Effect is weak and not yet replicated.

Themes: how richly to interpret infant behavior; active research area

Core knowledge of forms (shape skeleton)

Infants know about object shape; people assume shape tells you function (a fork's shape fits its use).

Shape skeleton: we extract an object's internal structure. When adults tap anywhere on a shape, taps cluster along the skeleton.

Ayzenberg and Lourenco 2022

Infant shape skeletons

Question: Are infants sensitive to shape skeletons?

Method: Habituation to one new object; test with objects that had the same skeleton (different surface parts) or a different skeleton.

Result: Infants looked longer at shapes with a DIFFERENT skeleton.

Claim: Infants represent overall shape by its skeleton and can form a shape category from one example.

Week 3 takeaway

Babies know a lot early: exact and approximate number, where they are, what agents are and that agents have goals, the structure of shapes.

But these abilities are fragile (signature limits), and it's unclear how important they are for later developme