Chapter 6: Generalization, Discrimination Learning, and Concept Formation

Core Concepts: Generalization, Discrimination, and Concept Formation

• Generalization: The transfer of past learning to novel events, stimuli, and problems.

• Discrimination Learning: The process by which animals or humans learn to respond differently to different stimuli.

• Concept Formation: The process by which individuals learn about new categories of entities in the world, usually based on shared or common features.

• Behavioral Processes Matrix:

  • Similar Stimuli $\rightarrow$ Same Outcome: Similar stimuli lead to identical predictions or consequences (e.g., both broccoli and cauliflower tasting nasty).

  • Similar Stimuli $\rightarrow$ Different Outcomes: Similar stimuli lead to distinct predictions or consequences (e.g., broccoli tasting nasty, but cauliflower tasting yummy).

  • Dissimilar Stimuli $\rightarrow$ Same Outcome: Distinct stimuli lead to identical predictions or consequences (e.g., both broccoli and red peppers tasting nasty).

  • Dissimilar Stimuli $\rightarrow$ Different Outcomes: Distinct stimuli lead to different predictions or consequences (e.g., broccoli tasting nasty, but red pepper tasting yummy).

Behavioral Processes of Generalization

• Generalization Gradient:

  • A curve showing how changes in the physical properties of stimuli (plotted on the horizontal axis) correspond to changes in responding (plotted on the vertical axis).

  • Demonstrates that an animal's response changes in a graded fashion based on the degree of similarity between a novel test stimulus and the original training stimulus.

  • Following reinforcement training with a single stimulus, the generalization gradient displays a peak (point of maximal responding) corresponding exactly to the original trained stimulus.

• Stimulus Generalization in Pigeons:

  • Pigeons trained to peck at a specific wavelength of light (e.g., yellow light) show maximum pecking rates to that specific wavelength.

  • Pecking rates decrease continuously as the light color shifts toward yellow-green or yellow-orange.

• Universal Law of Generalization and Consequential Region:

  • Consequential Region: A set of stimuli in the environment that share the same outcome or consequence as a stimulus whose consequence is already known.

  • Roger Shepard argued that the shape of generalization gradients reflects an animal's expectation that the likelihood of two stimuli having the same consequence drops off sharply as the stimuli become more physically distinct.

  • Generalization gradients represent an adaptive prediction attempt based on past experience to estimate whether novel stimuli share consequences with previously encountered stimuli.

Computational Models of Stimulus Representation

• Stimulus Representation: The formal scheme in which information about stimuli is encoded within a computational model or the brain.

• Discrete-Component Representation:

  • A representation scheme in which each individual stimulus (or stimulus feature) corresponds to a single individual node or element in the model.

  • Each input node connects via a modifiable weight to an output node.

  • Limitations: Fails to account for stimulus generalization among physically similar stimuli. If a model is trained on a yellow light, it shows zero response to a yellow-orange light because the input nodes are entirely independent. It produces an all-or-none response rather than a smooth generalization gradient.

• Distributed Representation:

  • A representation scheme in which information is coded as a pattern of activation distributed across many overlapping nodes or elements.

  • Physically similar stimuli activate overlapping subsets of input nodes.

  • Advantages: Naturally accounts for stimulus generalization. Training on one stimulus strengthens weights connected to shared nodes, causing similar stimuli to trigger partial responses and generating realistic, bell-shaped generalization gradients.

• Test Knowledge Application (Discrete vs. Distributed Representations):

  • Scenario A: A low-frequency tone predicts a shock, and a high-frequency tone predicts a shock, but a light predicts food. Predicting the response to a medium-frequency tone.

    • Model Required: Discrete-component representation is sufficient because all tone frequencies lead to the same shock outcome, removing the need to discriminate degree of similarity along the frequency spectrum.

  • Scenario B: Patients in a blue-walled hospital recover in $2\,\text{days}$, whereas patients in a red-walled hospital take $4\,\text{days}$. Predicting recovery time for patients in a green-walled hospital.

    • Model Required: Distributed representation is required because the model must capture the degree to which green overlaps in similarity with blue versus red.

Discrimination Learning and Stimulus Control

• Stimulus Control: The degree to which cues in the environment control or influence an organism's behavior.

• Herbert Jenkins' Discrimination Study:

  • Standard Training Group: Trained with $S (1000\,\text{Hz tone}) \rightarrow R (\text{key peck}) \rightarrow O (\text{food})$. Generalization test yields a broad curve centered at $1000\,\text{Hz}$.

  • Discrimination Training Group: Trained with $S (1000\,\text{Hz tone}) \rightarrow R (\text{key peck}) \rightarrow O (\text{food})$ alongside $S (950\,\text{Hz tone}) \rightarrow R (\text{key peck}) \rightarrow O (\text{no food})$.

  • Result: Discrimination training sharpens the generalization gradient, significantly narrowing the response curve around the positive stimulus ($S+$) and reducing responding to nearby frequencies.

• Dimensions of Discrimination:

  • Interdimensional Discrimination: Discrimination between stimuli that differ along a single continuous physical dimension (e.g., differentiating between two tone frequencies, such as $1000\,\text{Hz}$ vs. $950\,\text{Hz}$).

  • Extradimensional Discrimination: Discrimination between stimuli that differ across entirely distinct physical dimensions (e.g., differentiating between a tone and a light).

• Errorless Discrimination Learning:

  • A training procedure introduced by Herbert Terrace in which a difficult discrimination is acquired by starting with an easily distinguishable version of the task and incrementally progressing to harder variations as earlier stages are mastered.

  • Educational Applications: Highly effective in special education and for individuals with learning disabilities, leading to rapid, robust acquisition.

  • Trade-off: The resulting learned discrimination is extremely rigid and inflexible compared to standard trial-and-error learning.

Transfer Learning and Advanced Behavioral Paradigms

• Sensory Preconditioning:

  • A paradigm in which pre-exposure to two stimuli presented together as a compound leads to later transfer of conditioning from one stimulus to the other.

  • Experimental Structure:

    • Phase 1 (Compound Exposure): Experimental group receives compound stimulus (Tone + Light together); Control group receives separate presentations (Tone, Light separately).

    • Phase 2 (Conditioning): Both groups receive Light $\rightarrow$ Airpuff $\rightarrow$ Eyeblink response.

    • Phase 3 (Test): Tone alone is presented.

    • Results: Experimental group exhibits an eyeblink response to Tone alone; Control group exhibits no eyeblink response.

• Acquired Equivalence:

  • A generalization paradigm where prior training in stimulus equivalence increases generalization between two stimuli, even if those stimuli are superficially dissimilar.

  • Experimental Setup:

    • Phase 1 Training: $A1 \rightarrow X1 \rightarrow \text{food}$, $A2 \rightarrow X1 \rightarrow \text{food}$, $B1 \rightarrow Y1 \rightarrow \text{food}$, $B2 \rightarrow Y1 \rightarrow \text{food}$.

    • Phase 2 Training: $A1 \rightarrow \text{food}$, $B1 \rightarrow \text{no food}$.

    • Phase 3 Test: $A2$ evokes a strong pecking response (due to equivalence with $A1$); $B2$ evokes no strong pecking response.

• Negative Patterning:

  • A behavioral paradigm in which individual cues presented alone signal a positive response, but their combination (pattern) signals a negative response (no response).

  • Structure: Tone $\rightarrow$ Food, Light $\rightarrow$ Food, Tone + Light $\rightarrow$ No Food.

  • Computational Requirement: Single-layer network models using discrete-component representations cannot solve negative patterning because the combined activation of individual positive weights inevitably triggers the output. Solving it requires multi-layer networks with hidden configural units.

• Paradigm Identification Practice:

  • Scenario 1: Elizabeth likes men who bring candy or flowers on a first date, but is turned off if a man brings both.

    • Paradigm: Negative patterning (Candy $\rightarrow$ Good, Flowers $\rightarrow$ Good, Candy + Flowers $\rightarrow$ Bad).

  • Scenario 2: Samson learned in childhood that deep-voiced people have beards. Later he learns bearded men are strong, inferring deep voices also indicate strength.

    • Paradigm: Sensory preconditioning (Phase 1: Deep voice + Beard; Phase 2: Beard $\rightarrow$ Strong; Phase 3: Infer Deep voice $\rightarrow$ Strong).

  • Scenario 3: A music teacher plays snippets of Brahms, then Schubert, then Brahms to teach style recognition.

    • Paradigm: Discrimination training (Music 1 $\rightarrow$ Brahms, Music 2 $\rightarrow$ Schubert).

  • Scenario 4: Mark and Kaori like the same foods and people; Mark assumes Kaori will also like a song he enjoys.

    • Paradigm: Acquired equivalence (Mark $\rightarrow$ Food 1, Kaori $\rightarrow$ Food 1; Mark $\rightarrow$ Music 1 $\rightarrow$ Infer Kaori $\rightarrow$ Music 1).

• Variability in Training:

  • Training variability enhances skill generalization to novel tasks.

  • For learners with high initial knowledge/skill, high training variability produces superior generalization.

  • For learners with low initial knowledge, initial learning is maximized by starting with minimal variability before introducing variation.

Concept Formation, Category Learning, and Stereotypes

• Definitions:

  • Concept: An internal psychological representation of a real or abstract entity.

  • Category: A physical division or class of entities present in the world.

• Category Learning in Animals:

  • Pigeons possess complex discrimination abilities, capable of categorizing abstract vs. impressionist paintings, as well as baroque vs. neoclassical music styles.

• Prototypes and Natural Categories:

  • Eleanor Rosch demonstrated that natural categories lack rigid boundaries; members vary along a spectrum of typicality.

  • Prototype: The central tendency or idealized average representation of a category.

  • Family Resemblance: Categories structured around overlapping clusters of shared features permit inductive inference (drawing probable general rules or predictions from specific instances).

• Stereotypes, Discrimination, and Social Generalization:

  • Stereotype: A set of psychological beliefs regarding the attributes of members of a specific social group.

  • Social Discrimination: Unfair differential treatment of individuals based on their perceived group membership.

  • Acquisition & Bias: Stereotypes are acquired through personal filters based on self-interest and needs. They are sustained via confirmation bias—the tendency to selectively attend to and remember information that confirms pre-existing stereotypes.

  • Stereotype vs. Prototype: Prototypes are defined externally by true central tendencies of category exemplars. Stereotypes are subjective psychological concepts that often fail to accurately mirror real-world prototypes.

  • Appropriate vs. Inappropriate Use: Using generalizations appropriately requires balancing specificity (does it apply exclusively to this group?) and generality (does it apply to all members?). Inappropriate use occurs when applying non-statistical generalizations or using group-level stats to deny an individual's unique variations.

Neural Substrates of Generalization and Discrimination

• Cortical Representations of Sensory Cues:

  • Cortical space is allocated non-uniformly based on behavioral importance (e.g., a New Yorker's mental map exaggerates local avenues while shrinking the entire Midwest).

  • Primary Auditory Cortex ($A1$): Organized tonotopically; neighboring neurons respond preferentially to adjacent sound frequencies.

  • Shared-Elements Cortical Models: Physically similar stimuli activate overlapping populations of cortical neurons.

• Role of Primary Sensory Cortex in Generalization:

  • Richard Thompson demonstrated that an intact primary sensory cortex ($A1$) is essential for normal, fine-grained stimulus generalization.

  • Without $A1$, animals can still learn to detect the crude presence of a tone, but lose tone-specific response curves and accurate generalization.

• Cortical Plasticity and Remapping:

  • Norman Weinberger showed that pairing a $2500\,\text{Hz}$ tone with a shock alters $A1$ tuning: neurons previously tuned to $1000\,\text{Hz}$ shift their peak responsiveness toward the $2500\,\text{Hz}$ training frequency.

  • Primary sensory cortices ($A1$, $V1$, $S1$) reallocate cortical space based on cue salience.

• Nucleus Basalis and Acetylcholine:

  • Nucleus Basalis: A nucleus in the basal forebrain that projects cholinergic axons throughout the cortex.

  • Releases acetylcholine (ACh), enabling synaptic plasticity and cortical remapping.

  • Lesion Effects: Surgical destruction of cholinergic neurons in the nucleus basalis does not impair discrimination habits acquired before surgery, but completely blocks the acquisition of new discrimination learning post-surgery due to the loss of cortical plasticity.

The Hippocampal Region in Generalization and Clinical Conditions

• Anatomy of the Hippocampal Region:

  • Comprises the hippocampus, entorhinal cortex, dentate gyrus, and subiculum (referred to as the medial temporal lobe / MTL in humans).

• Role in Contextual and Relational Learning:

  • Surgically severing the fornix (a main hippocampal tract) in rabbits completely abolishes sensory preconditioning.

  • Latent Inhibition: A phenomenon where pre-exposure to a CS without a US retards subsequent CS-US conditioning. Damage to the hippocampal region eliminates latent inhibition.

  • The hippocampal region is critical for stimulus generalization tasks that depend on learning relational associations between stimuli (sensory preconditioning, acquired equivalence, latent inhibition).

• Gluck-Myers Model of Hippocampal Function:

  • Proposes that the hippocampal region acts as an "information gateway" during learning.

  • It compresses redundant information and creates new representations that dictate how other cortical regions permanently store and map stimulus relationships.

• Generalization Deficits in Schizophrenia:

  • Schizophrenia involves structural abnormalities in hippocampal morphology and reduced hippocampal activation.

  • Patients display severe impairments in acquired-equivalence tasks, transitive inference, and chaining tasks.

  • Shohamy and Wagner demonstrated via fMRI that hippocampal activation during initial learning directly correlates with individual accuracy during subsequent generalization testing.

• Altered Generalization in Autism Spectrum Disorder (ASD):

  • ASD is linked to variations in neural connectivity across cortical and subcortical networks.

  • Stimulus Overselectivity: Children with ASD struggle to process compound cues, failing to integrate multiple sources of information and instead binding rigidly to a single isolated element.

  • Degree of overselectivity correlates directly with clinical severity in social impairment, language deficits, poor inhibitory control, and ritualistic behaviors.

Practical Application & Rescorla Concept Check

• Robert Rescorla's Experiment (1976):

  • Phase 1: All rats are trained on Yellow Light CS $\rightarrow$ US until reaching criterion CR performance.

  • Phase 2: Experimental group receives additional training with Orange Light CS $\rightarrow$ US; Control group continues training exclusively with Yellow Light CS.

  • Phase 3: Both groups are tested with Yellow Light CS.

• Predictions & Theoretical Principles:

  • Phase 3 Outcome: Rats in the experimental group display a smaller response to the yellow light than control rats.

  • Mechanism: Phase 2 training with the orange light forces the experimental rats to narrow their generalization gradient (learning a sharper stimulus representation between closely adjacent visual wavelengths), which reduces response generalizeability back to the yellow light.

  • Application to Athletic Performance (Tennis): To generalize skills effectively to a tournament setting, an athlete should train under diverse, variable contexts (varying court surfaces, lighting, opponent styles) rather than repeating highly narrow, uniform drills. Training variability broadens adaptive stimulus representations, facilitating transfer to novel competitive scenarios.