PSY260 - 5. Generalization and Discrimination Learning

Question 1

Generalization

Answer 1

transfer of past learning to new situations + problems

Question 2

Specificity + Generality

Answer 2

deciding how narrowly rule applies + how broadly rule applies

Question 3

Discrimination

Answer 3

perception of diff betw stimuli

Question 4

Generalization gradient

Answer 4

curve showing how changes + physical properties of stimuli correspond to changes in responding
•After training in which single stimulus has been reinforced repeatedly, generalization gradients around that turn stimulus form peak, corresponding to original stimulus on which animal was trained
•Declined rapidly on either side of peak

Question 5

Generalization as search for similar consequences

Answer 5

• Challenge of generalization is to identify consequential region: set of all stimuli that have same consequence as training stimulant
• consistently expect chance 2 stimuli will have same consequence drops off sharply as stimuli become more distinct

Question 6

Distributed representations

Answer 6

stimuli represented by overlapping sets of nodes
Similarity emerges naturally from fact that 2 similar stimuli activate elements belonging to both sets
•What is learned about one stimulus will tend to transfer to other stimuli that activate some of same nodes

Question 7

Shared elements and distributed representations

Answer 7

• Three layers of nodes + 2 layers of weight: 2 layer network
• Each stimulus activate + input node is connected to several nodes in internal representation by fixed weights
• Internal representation nodes connected via modifiable weight, which will change during learning, to final output node

Question 8

Topographic representation

Answer 8

nodes representing physically similar stimuli placed next to each other in model

Question 9

Shared elements and distributed representations

Answer 9

• Weights from 3 active nodes = 0.33
• Yellow light activates nodes 3,4 + five = 1.0 in output node
• Width + shape of generalization gradient produced by model can be manipulated by varying # + amount of overlap in nodes in model
• Distributed representation systems capture fundamental property of learning: organisms tend to treat similar event similarly and to expect similar stimuli to have similar consequences

Question 10

Peak shifts in generalization

Answer 10

• Pigeons were trained to pack at 550 nm light
• Control group S+ received only this training + S minus given additional discrimination trials with very similar light at 555 nm that was not reinforced
• Peak shift: Peak responding of experimental animal shifted away from non-reinforced stimulus

Question 11

Peak shifts in generalization

Answer 11

• Positive generalization gradient develops around S+ + inhibitory gradient develops around S-
• If net associative strength for any to stimulus is diff betw excitatory gradient formed around S+ + inhibitory gradient formed around S– + if these 2 gradients overlap considerably, then greatest net positive strength will be to the left of S+ because region is both close to S+ [550] + farther away from S- [555]

Question 12

Peak shift and the Rescorla-Wagner Model in Fixed Ration operant conditioning

Answer 12

• Superconditioning – emerges from weight summation + cue competition property of Rescorla-Wagner model
• The trough + peek in model results depict responses that super conditioned to be even lower/higher than other elements on same side of the transition point

Question 13

Discrimination training with multiple dimensions: implications for attention allocation

Answer 13

• People trained on 1 dimension [size/brightness] were subsequently more able to make distinctions between stimuli that differ along that dimension than betw stimuli that differ along at diff, untrained dimension
• Generalization gradients can be modified by experience so organism pays attention to whatever most relevant stimulus dimensions in a given situation

Question 14

Negative patterning: differentiating configural cues from their individual components

Answer 14

• cue combinations have radically diff meanings than their components (left signal + right signal = hazard)
• Negative patterning: response to individual cues is positive while response to compound is negative
• Difficult to learn because it requires surprising natural tendency to generalize about similar stimuli
• Configural node: detector for unique configuration of two cues. Will fire only if all inputs are active
• When configural node is activated, it cancels out effects of tone only + light only nodes, for net input of 0

Question 15

Configural learning in categorization

Answer 15

• Category learning: process by which animals + humans learn to classify stimuli into different categories
• Combinatorial explosion: stems from rapid expansion of resources required to encode configurations as number of component features increases

Question 16

Configural learning in categorization

Answer 16

• Network has 2 layers of modifiable weights – one going from input nodes to internal nodes + another going from internal nodes to output node
• When experience shows particular combo of stimuli to be useful for purposes of solving problems, changes in values of lower layer weights can cause network adapter that in future, particular internal node will become active only went that specific combo of inputs is present

Question 17

Sensory preconditioning: cooccurrence and stimulus generalization

Answer 17

•Sensory preconditioning: Coocurrents of 2 stimuli can lead to generalization, prior presentation of two stimuli together as compound result in later tendency for any learning about one of the stimuli to generalize to the other
•Meaning based generalization: tone + light have same meaning even though they do not have any relevant physical similarity
Similarity based generalization: arises naturally between two stimuli that are physically similar
•Cooccurrence of 2 stimuli is sufficient to produce meaning based generalization from one stimulus to another

Question 18

Acquired equivalents: novel similar predictions based on prior similar consequences

Answer 18

• Meaning based generalization can occur wen 2 noncombined stimuli share the same consequence
• Can learn to generalize from one stimulus to another that is superficially similar if stimuli have a history of cooccuring or of predicting same consequence

Question 19

Probabilistic category learning and transfer generalization

Answer 19

• Ppl very poor at making kinds of probability estimates one base rates for two categories are very different, failure in decision-making known as base rate neglect
• Accuracy of peoples info about category maybe very tightly bound to particular way they learn properties of that category
• Learning may not generalize well if subsequent transfer task challenges learners to apply knowledge in novel ways

Question 20

Decision making errors and misapplications of generalization

Answer 20

• Common misuse of generalization comes from faulty inverse reasoning about categories
• NBA players are black and Black people are NBA players

Question 21

Cortical representations and generalization

Answer 21

• When perceptual learning occurs in humans, cortical changes company enhanced discrimination abilities
• Selectivity of individual cortical neurons in responding to specific stimulus features can be modified through experience
• Learning can change spatial organization of interconnected neurons and sensory cortex

Question 22

Cortical representations of sensory stimuli

Answer 22

• Homunculus: parts of the body that are especially sensitive to touch activate larger areas of S1
• Primary somatosensory shows similar organization with homunculus replaced by distorted figure of species in question, altered to reflect body areas important to animal
• Receptive field for neuron: range of physical stimuli that activate it
• Wider neuron’s receptive field is, brother the range of physical stimuli that will activate neuron

Question 23

Shared-elements models of receptive fields

Answer 23

• Physically similar stimuli will activate, notes, or neurons
• To similar tones should cause overlapping sets of neurons to fire
• Simplified network makes plan why cortical neurons display receptive fields
• Neurons best frequency is 550 Hz; similar tones also activate this neuron, although not as strongly as tone of 550 Hz
• Result in generalization gradient that looks like actual receptive fields obtained during cortical mapping studies

Question 24

Topographic organization and generalization

Answer 24

• Although it is possible for an animal to learn to respond to stimuli while lacking the corresponding areas of sensory cortex, and intact sensory cortex for that stimulus type is essential for normal generalization
• Without A1, animals can learn to respond to the presence of a tone but cannot respond precisely to a specific tone
• Without primary sensory cortex, animals overgeneralize and have difficulty discriminating stimuli and corresponding sensory modality

Question 25

Plasticity of cortical representations

Answer 25

• Lack of stimulation or use can cause shrinking of cortical areas
• Lost lend results in region of cortex to remain idle, nearby areas of homunculus may spread into vacated space
• Areas Acquire increased cortical representation and consequent increased sensitivity to stimulation in touch
• Cortical plasticity is a result of the tone shock pairing
• Stimulus presentation alone doesn’t drive cortical plasticity; stimulus has to be meaningfully related to ensuing consequences, such as a shock

Question 26

Plasticity of cortical representations

Answer 26

• If cortical change occurs because stimulus in one sensory modality is meaningfully related to or predictive of a salient consequence in a different sensory modality, how did info about consequence reach primary sensory cortex of first modality in order to produce change there
• Primary sensory cortex of first modality only receives info that some sort of salient event has occurred
• info is enough to instigate cortical remapping and expand representation of the cue stimulus

Question 27

Plasticity of cortical representations

Answer 27

•Primary sensory Cortices is only determine which stimuli deserve expanded representation within that primary cortex and which do not
•Basal forebrain group of nuclei important for learning and memory
oDamage can produce anterograde amnesia – severe impairment in forming in fact an event memories

Question 28

• Nucleus basalus

Answer 28

oWhen activated, the release acetylcholine a neurotransmitter that promotes neuronal plasticity
oWhen CS is paired with the US, nuclear this basalis becomes active and delivers acetylcholine to cortex, enabling cortical remapping enlarge representation of that CS
oIt receives info through connections from areas such as the amygdala, which codes emotional info such as discomfort and pain and pleasure

Question 29

Effect of damage to the hippocampal region

Answer 29

• Rabbits with lesions display no sensory preconditioning
• Lesion studies show that hippocampal region is critical for meaning based generalization
• Latent inhibition is eliminated by hippocampal region damage
• Hippocampal region lesions and animals – sensory preconditioning, acquired equivalents, and blatant inhibition – suggest brain region is critical for a stimulus generalization, especially when generalization involves learning relationships between different stimuli

Question 30

Modeling role of Hippocampus in adaptive representations

Answer 30

• Hippocampal region appears to be critically involved in developing new representations
• Operates as an information Gateway during associative learning, storing new representations of events that are experienced
• Selects what info was allowed to enter memory and how it is encoded by other brain regions
• Unimportant or redundant info undergoes shrinkage, useful info is expanded or differentiated, creating new, efficient, optimize representation that includes only key aspects of incoming info