Dev Psych Week 4: Domain-General Learning (Concepts)

Learning

Definition: a change in behavior or knowledge that lasts over time and is caused by experience

Not learning:

Maturation: lasting changes programmed by genes (puberty)

Transient changes: caused by the environment but temporary (most drug effects)

Framing question: if you built a simulated infant, which learning mechanisms would you build in?

Associative learning

Definition: the most basic form of learning; things that regularly happen together come to trigger each other. Found in infants, adults, animals, simple organisms

Hebb's rule: "neurons that fire together, wire together"

Types:

Classical conditioning: neutral stimulus paired with an involuntary response until the stimulus alone triggers it (Pavlov). Newborns can do it but need a longer delay (~1.5 s vs. ~600 ms) = slower processing

Operant learning: voluntary behavior linked to a consequence

Operant (reinforcement) learning

Definition: a voluntary behavior is linked to a consequence

Reinforcement: makes behavior more frequent

Punishment: makes behavior less frequent

Conjugate reinforcement: reward changes immediately and in proportion to the behavior (kick harder, mobile moves more)

Phases: baseline -> acquisition (behavior rises) -> extinction (drops) -> reacquisition (faster = memory)

Infant memory: 6-month-olds remember a contingency ~2 weeks; older longer, younger less

Punishment

Negative punishment: remove something good (voice stops, time-out); most studied because it's ethical

Positive punishment: add something bad (bitter liquid)

Problem: young infants struggle to stop a behavior that used to be rewarded

Credit assignment problem: punishment says what's wrong but not what to do; the child can't tell what caused it (the table? the toys? the parent?)

Real world: reinforce the desired behavior instead; if needed, use time-out; spanking is ineffective and likely harmful

Statistical learning

Definition: extracting regularities (patterns) from the environment

Transition probability: how likely one item follows another; high within words, low across word boundaries, so rare transitions mark boundaries

vs. associative: no reward needed, works with weak patterns (~20% vs ~90%), largely implicit

Key ideas:

Present from birth, even in sleeping newborns

Infant hippocampus tracks it like adults'

Infants match or beat adults (development ≠ improvement)

Used for word segmentation and word learning

Error-driven learning

Definition: trial and error; act, notice the error, use its direction and size to adjust, try again

Example: force-field reaching; a hidden force pushes your hand, and by ~100 trials you aim off-center, possibly without awareness

Why it's efficient: like video compression, store the first frame then only the differences (errors)

Surprise and exploration

Dark room problem: if we only minimized error, we'd hide in a dark room and never learn; so we also seek surprise

Goldilocks effect: infants look away from events that are too predictable or too surprising; attend most to moderate information (U-shaped curve). Link to habituation: repeated stimuli lose surprise

Children as little scientists: after a violated expectation, infants test that specific expectation (bang a toy that went through a wall, drop one that floated); surprise boosts exploration and learning

Prediction

Predictive coding: the brain encodes only prediction errors, the difference between what happened and what it expected

Anticipatory looking: after learning a sequence, infants look to where the next item will appear before it appears

Omission trials: when an expected stimulus doesn't appear, the infant brain (e.g., visual cortex) still activates, showing it predicted it

Other kinds of learning

Social learning: learning from other people (imitation); covered later

Perceptual learning: long-lasting, experience-dependent improvement in perception (wine experts, birdwatchers); change is in the brain, not the sense organ; doesn't depend on strategy

Examples:

Childhood experience shapes visual cortex specialization (Pokemon region in childhood players)

Oblique effect: better detection near vertical/horizontal than diagonal, because the world has more of those lines; perception is tuned to the statistics of what we see

Capacities infants probably don't have

Imprinting: attaching to whatever is seen right after birth in a critical period (ducklings); related to genetically triggered learning (songbirds); no evidence in humans

One-shot learning: remembering after a single exposure; needed for episodic memory; appears ~18 months

Rule-based learning: extracting abstract patterns (ABA) that generalize; original evidence failed to replicate (theme: active research area)

Backpropagation: main AI algorithm, impractical for brains (every new input would update the whole network)

Children as scientists: types of reasoning

Children build theories, test them, and revise them

Inductive: sample -> population and back | Limit: need more samples than adults

Causal: notice reversed cause and effect | Limit: assume causes need contact

Deductive: possibly disjunctive syllogism at 12 months | Limit: interpretation debated

Counterfactual: little or none | Limit: develops slowly

Inductive inference

Definition: reasoning from a sample to a population (and back); faster than statistical learning, which needs a lot of input

Examples:

Infants infer an object's 3D shape from a few views; before ~5 months they need several views (chicks need views in continuous motion)

8-month-olds expect a box's contents to match a sample drawn from it, in both directions

Causal reasoning

Hume: we never see causation directly, only constant conjunction (things happening together); a one-directional, perfect correlation is perceived as causation

Key ideas:

Launching: 6-month-olds notice when cause and effect are reversed (direct launch looks causal, delayed doesn't)

Contact: kids assume causes need contact; evidence can override it

Predict ≠ intervene: toddlers predict an effect before they can act to cause it

Failure: 16-month-olds infer whether a failure is their fault or the toy's

Deductive inference

Definition: reasoning with certainty from premises to a conclusion

Disjunctive syllogism: A or B; not A; therefore B (reasoning by elimination)

Ex: either there's a long line at the cafe or it's closed; it isn't closed; so there's a long line

Infants: 12-month-olds may do this (looking + pupil dilation), but interpretation is debated (theme: how richly to interpret infant behavior)

Counterfactual reasoning

Definition: imagining a different outcome: "if A had not happened, would B?"

Debate: 3-4-year-olds seem to answer "what if" questions correctly, but with tighter tests true counterfactual reasoning only appears around 12-14 years; younger kids may use simpler rules

Takeaway: infants probably can't do it; it develops slowly

Face perception: domain-specific vs. domain-general

Domain-specific view: infants are born with a face template and face-specialized brain regions; experience refines it

Evidence: newborns track face-like patterns; prefer upright face-like patterns with dark eye/mouth regions and direct gaze; fusiform face area (even in blind adults touching faces)

Domain-general view: general visual biases give only weak face preference; experience is critical

Evidence: poor newborn acuity; mother preference disappears with headscarves (hairline) and needs her voice; face-deprived monkeys develop no face areas

Theme: nature and nurture

Perceptual narrowing (faces)

Definition: infants lose sensitivity to what they don't experience and get better at what they do

Faces: 6-month-olds tell apart both human and monkey faces; 9-month-olds and adults only human faces

Other-race effect: infants come to discriminate own-race faces better than other-race faces

Escaping it: the right experience (naming individual monkey faces) prevents narrowing

Theme: development does not equate to improvement; reconciles domain-specific and domain-general views