Generalization: transfer of past learning to novel events. The brain's default — treat similar things similarly.
Discrimination: perception of differences between stimuli; knowing when not to generalize.
Core tension: Too broad = fails (broccoli = all vegetables). Too narrow = fails (yellow = only exactly 580 nm). Adaptive behavior requires calibrated generalization.
| Feature | Discrete-Component | Distributed (Shared Elements) |
|---|---|---|
| Encoding | One unique node per stimulus | Overlapping node sets for similar stimuli |
| Generalization | None — untrained node = response 0 | Automatic — trained nodes activate for similar stimuli |
| Example (yellow → orange) | Orange node weight = 0; response = 0 | Node 5 shared → response = 0.33 |
| When sufficient | Categorically distinct stimuli (tone vs. light); simple organisms | Stimuli on continuous dimensions (color, pitch, brightness) |
Yellow-orange → nodes [4, 5, 6] → response = 0.33+0.33 = 0.66
Orange → nodes [5, 6, 7] → response = 0.33 (only node 5 trained)
A1 (primary auditory cortex) is tonotopically organized. Tuning curves get broader as signals climb from cochlea to cortex — more generalization at higher levels, not less. Removing A1 in cats eliminates gradient (massive overgeneralization). A1 is necessary for appropriate generalization.
Discrimination training sharpens gradient (Jenkins & Harrison, 1962): Standard 1000 Hz training → broad gradient. Discrimination (1000 Hz food; 950 Hz no food) → much steeper, narrow gradient centered on 1000 Hz. Why: S- builds inhibitory gradient counteracting generalization from S+.
Human analog (Wills & McLaren, 1997): A-only training → broader gradients than A-vs-B discrimination training.
- Universality: Peak shift occurs in bees, horses, rats, goldfish, chickens, pigeons, and humans — one of the most robust findings in learning research.
- Human relevance: Caricatures work via peak shift — exaggerating features that distinguish a face from the average face.
Not all generalization is similarity-based. Two stimuli sharing the same consequences can generalize even without physical similarity. Three paradigms illustrate this — know all three cold.
4.1 Sensory Preconditioning (3-phase procedure)
Phase 2: Light → Airpuff [both groups learn to blink to light]
Phase 3: Test Tone alone → compound group blinks; control does NOT
Interpretation: Co-occurrence creates association. Light's meaning (airpuff) transfers to tone — cross-modality, meaning-based generalization without physical similarity.
Hippocampal dependence: Rabbits with fornix lesions show NO sensory preconditioning. Lesioned compound-exposure animals look exactly like controls (Port & Patterson, 1984).
4.2 Acquired Equivalence
A1 and A2 both paired with X1 → become "equivalent." Then A1 reinforced → animals generalize strongly to A2 but not to B2 (different equivalence class).
Human version (Myers et al., 2003): Cartoon people preferring fish. Brown-haired and blonde both prefer blue fish. Brown-haired newly prefers red fish → adults infer blonde also prefers red fish.
Hippocampal dependence: Entorhinal cortex lesions in rats eliminate Phase 3 transfer. Phase 1+2 intact but generalization fails.
4.3 Latent Inhibition
Pre-exposing animal to CS (without US) slows later CS-US conditioning. Must "unlearn" tone-nothing association before building tone-shock.
Entorhinal cortex lesions ELIMINATE latent inhibition — lesioned pre-exposed animals learn as fast as controls. Brain damage HELPS learning speed by removing the prior inhibitory memory.
Weinberger's studies: Paired 2500 Hz tone with shock in guinea pigs. A1 neurons that originally preferred ~1000 Hz shifted to respond best to ~2500 Hz (trained frequency). More A1 neurons become tuned to trained stimulus. Requires pairing (CS-US): tone alone → habituation; unpaired → no shift.
How does A1 "know" what's important? A1 doesn't directly receive input from pain/reward systems. The answer: Nucleus Basalis.
| Neuromodulator | Source | Signal | Role |
|---|---|---|---|
| Dopamine | VTA / Substantia Nigra | Valence (signed PE: good or bad) | Reward/punishment prediction error |
| Acetylcholine (ACh) | Nucleus Basalis (basal forebrain) | Salience (unsigned: how important) | Cortical plasticity & remapping |
→ ACh released broadly across cortex
→ Enables cortical remapping (tuning shifts toward trained stimulus)
- Nucleus basalis lesion + prior auditory discrimination: Previously learned auditory discrimination intact (remapping already happened).
- Nucleus basalis lesion + new visual discrimination: IMPAIRED — cannot narrow gradient to distinguish colors; overgeneralizes. New cortical remapping requires ACh.
The hippocampus is NOT required for simple S-R associations. It IS required for meaning-based generalization (sensory preconditioning, acquired equivalence) and context-dependent discrimination (latent inhibition).
Gluck & Myers (1993) model: Hippocampus compresses redundant/unimportant information AND differentiates (expands the representation of) useful information.
New Yorker map analogy: Fine detail on 9th/10th Ave (expanded for what matters); featureless Midwest (compressed for what doesn't). This is what the hippocampus does with stimulus representations.
Prototype: "Average" or most central member of a category. Assign new observations by distance from prototype. Animals can form categories: pigeons distinguish benign/malignant tumors at radiologist-level accuracy.
Sapir-Whorf hypothesis: The language you speak can affect how you categorize the world.
| Example | Observation | Principle |
|---|---|---|
| Russian blue shades | Two words for blue → Russian speakers show finer blue discrimination than English speakers (same visual hardware) | Category labels tune generalization gradients |
| Arabic camel (~160 words) | Finer discrimination of camel types | Rich vocabulary → more cortical resources → finer gradient |
| English hen/rooster vs. one word for lion | Gender-specific terms improve discrimination | Naming creates category boundary |
Neural substrate: Schizophrenia is associated with reduced hippocampal volume and reduced hippocampal fMRI activity.
Core behavioral profile: Learn simple associations NORMALLY; TRANSFER generalization impaired.
| Task | Performance | Interpretation |
|---|---|---|
| Acquired equivalence — Phases 1 & 2 (acquisition) | Normal | Basic associative learning intact |
| Acquired equivalence — Phase 3 (transfer) | Significantly impaired (Keri et al., 2005) | Hippocampal encoding of equivalences failed |
| Transitive inference — non-relational pairs (AE) | Succeed | Rote memory intact (A always wins, E always loses) |
| Transitive inference — relational pairs (BD) | Fail | Requires hippocampal relational inference across chain |
- Antipsychotic medications partially remediate the transfer generalization deficit.
- Behavioral signature: Intact acquisition + impaired generalization transfer points specifically to hippocampal dysfunction.
Key Terms — 20 Flashcards
Click any card to reveal its definition. Use the filters to focus on a category.
Big Picture Synthesis
How the pieces connect — the unifying framework for Module 9.
How a stimulus is REPRESENTED in the brain determines everything about what generalizes to what. The same learning mechanism (prediction error update) produces radically different generalization patterns depending on whether stimuli are encoded discretely or as overlapping distributed patterns.
The hippocampus does not learn associations directly — it builds the representational scaffold (compressing irrelevant variation, differentiating useful distinctions) that lets other brain areas generalize correctly.
- Discrete vs. distributed — which model is required when (hospital color example)
- Peak shift direction — away from S-, know S+ and S-
- 3-phase structures — sensory preconditioning and acquired equivalence
- Hippocampus deficit — Phase 3 ONLY, not acquisition
- ACh vs. dopamine — salience vs. valence
- Sapir-Whorf — directionality debate unresolved
- Schizophrenia — intact acquisition + impaired transfer
- R-W (Weeks 5-8): Distributed representation implements graded generalization R-W describes mathematically
- ACh + Dopamine: Complementary neuromodulators — salience vs. valence
- Hippocampus (Week 6): Same structure as spatial learning and declarative memory
- Schizophrenia → Week 10: Clinical generalization deficits connect to psychiatric disorders module
- Cortical maps: Same principle as homunculus, violinists' hand expansion, and Weinberger's remapping
- Peak shift direction: Peak moves AWAY from S- (not toward it). S+=550, S-=555 → peak at ~540, NOT ~565.
- Latent inhibition lesion paradox: Entorhinal lesion HELPS speed (not hurts) — removes interference from prior tone-nothing memory.
- Hippocampus timing: Critical at ENCODING (Phase 1), not retrieval (Phase 3). fMRI at Phase 1 predicts success; Phase 3 activity does not.
- ACh vs. dopamine: ACh = unsigned salience (good or bad matters equally). Dopamine = signed valence (which direction matters).
- Nucleus basalis lesion: Old discriminations INTACT; only NEW discriminations fail. ACh needed for future remapping, not maintenance of past.