Development, Disintegration, and Neural Basis of Semantic Cognition: A Parallel-Distributed Processing Approach James L. McClelland Department of Psychology and Center for Mind, Brain, and Computation Stanford University In collaboration with: Tim Rogers, Karalyn Patterson, Matt Lambon Ralph, and Katia Dilkina

Representation is a pattern of activation distributed over neurons within and across brain areas. Bidirectional propagation of activation underlies the ability to bring these representations to mind from given inputs. The knowledge underlying propagation of activation is in the connections. Experience affects our knowledge representations through a gradual connection adjustment process language Parallel Distributed Processing Approach to Semantic Cognition

A Principle of Learning and Representation in PDP Networks Learning and representation are sensitive to coherent covariation of properties across experiences.

What is Coherent Covariation? The tendency of properties of objects to co- occur in clusters. e.g. –Has wings –Can fly –Is light Or –Has roots –Has rigid cell walls –Can grow tall

Development and Degeneration Sensitivity to coherent covariation in an appropriately structured Parallel Distributed Processing system underlies the development of conceptual knowledge. Gradual degradation of the representations constructed through this developmental process underlies the pattern of semantic disintegration seen in semantic dementia.

Some Phenomena in Development Progressive differentiation of concepts Overgeneralization Illusory correlations

The Rumelhart Model

The Training Data: All propositions true of items at the bottom level of the tree, e.g.: Robin can {grow, move, fly}

Target output for ‘robin can’ input

ajaj aiai w ij net i =  a j w ij w ki Forward Propagation of Activation

 k ~ (t k -a k ) w ij  i ~   k w ki w ki ajaj Back Propagation of Error () Error-correcting learning: At the output layer:w ki =  k a i At the prior layer: w ij =  j a j … aiai

ExperienceExperience Early Later Later Still

Waves of differentiation reflect coherent covariation of properties across items. Patterns of coherent covariation are reflected in the principal components of the property covariance matrix. Figure shows attribute loadings on the first three principal components: –1. Plants vs. animals –2. Birds vs. fish –3. Trees vs. flowers Same color = features covary in component Diff color = anti-covarying features What Drives Progressive Differentiation?

Overgeneralization of Frequent Names to Similar Objects “dog” “goat” “tree”

Illusory Correlations Rochel Gelman found that children think that all animals have feet. –Even animals that look like small furry balls and don’t seem to have any feet at all. A tendency to over-generalize properties typical of a superordinate category at an intermediate point in development is characteristic of the PDP network.

A typical property that a particular object lacks e.g., pine has leaves An infrequent, atypical property

Sensitivity to Coherence Requires Convergence A A A

Development and Degeneration Sensitivity to coherent covariation in an appropriately structured Parallel Distributed Processing system underlies the development of conceptual knowledge. Gradual degradation of the representations constructed through this developmental process underlies the pattern of semantic disintegration seen in semantic dementia.

Disintegration of Conceptual Knowledge in Semantic Dementia Progressive loss of specific knowledge of concepts, including their names, with preservation of general information Overgeneralization of frequent names Illusory correlations

Picture naming and drawing in Sem. Demantia

Grounding the Model in What we Know About The Organization of Semantic Knowledge in The Brain Evidence from Martin and others indicates that specialized brain areas subserve many different kinds of semantic information. Semantic dementia results from progressive bilateral disintegration of the anterior temporal cortex. Rapid acquisition of new knowledge depends on medial temporal lobes, leaving long-term semantic knowledge intact. language

Proposed Architecture for the Organization of Semantic Memory color form motion action valance Temporal pole name Medial Temporal Lobe

Rogers et al (2005) model of semantic dementia Gradually learns through exposure to input patterns derived from norming studies. Representations in the temporal pole are acquired through the course of learning. After learning, the network can activate each other type of information from name or visual input. Representations undergo progressive differentiation as learning progresses. Damage to units within the temporal pole leads to the pattern of deficits seen in semantic dementia. nameassocfunction temporal pole vision

Severity of DementiaFraction of Neurons Destroyed omissionswithin categ. superord. Patient Data Simulation Results Errors in Naming for As a Function of Severity

Simulation of Delayed Copying Visual input is presented, then removed. After several time steps, pattern is compared to the pattern that was presented initially. Omissions and intrusions are scored for typicality nameassocfunction temporal pole vision

Omissions by feature typeIntrusions by feature type IF’s ‘camel’ DC’s ‘swan’ Simulation results

Development and Degeneration Sensitivity to coherent covariation in an appropriately structured Parallel Distributed Processing system underlies the development of conceptual knowledge. Gradual degradation of the representations constructed through this developmental process underlies the pattern of semantic disintegration seen in semantic dementia.

Further Investigations Lexical and Semantic Deficits in SD Individual Differences Laterality Effects Unilateral Resections and Recovery Effects of TMS on Semantic Task Performance Category specific deficits: –Why they may occur in some types of patients but not others.

Further Investigations  Lexical and Semantic Deficits in SD Individual Differences Laterality Effects Unilateral Resections and Recovery Effects of TMS on Semantic Task Performance  Category specific deficits:  Why they may occur in some types of patients but not others.

Relation between lexical and semantic deficits (Patterson et al, 2006) Tested 14 SD patients, each assigned a ‘Semantic Score’ based on 3 standard tests. Then tested each patient on: –Reading HF&LF Reg. and Exc. Words –Spelling HF&LF Reg. and Exc. Words –Past Tense Inflection, HF&LF R&E Words –Lexical Decision: fruit/frute, flute/fluit –Object Decision (at right) –Delayed Copying Test

Reg. Exc. Reg. Exc.

The Integrated Connectionist Model

Category Specific Deficits: What is the cause? SD patients typically do not show a differential deficit for living things; however, Herpes Encephalitis patients often do. Alternative explanations? –Differential lesion location –Differential nature of the lesion itself In the Rogers Et Al model, removing links produces the SD pattern, but adding noise to connection weights produces the Herpes pattern (Lambon-Ralph, Lowe, and Rogers, 2007).

Naming of Animals and Artifacts at the Basic Level: Patients and Simulation Results From Lambon-Ralph, Lowe, & Rogers 2007

Predictions from Model Herpes patients should make relatively more semantic errors, SD patients relatively more omissions. When tested at the subordinate level of naming, differences between Herpes and SD patients should go away.

Conclusions A single set of principles provide an integrated framework for developing an integrated understanding of both the development and the neural basis of semantic cognitive abilities. For neuropsychology, the models capture the graceful degredation of semantic knowledge with brain damage, as well as frequency, typicality, and even some aspects of apparent domain-specificity effects. There are several promising avenues for further extension of the model.

The Integrated Connectionist Model

Sensitivity to Coherence Requires Convergence A A A

Coherence Training Patterns No labels are provided Each item and each property occurs with equal frequency Properties Coherent Incoherent Items is can has is can has …

Effect of Coherence on Representation

Another key property of the model Sensitivity to coherent covariation can be domain- and property-type specific, and such sensitivity is acquired as differentiation occurs. Obviates the need for initial domain-specific biases to account for domain-specific patterns of generalization and inference.

Differential Importance (Marcario, 1991) 3-4 yr old children see a puppet and are told he likes to eat, or play with, a certain object (e.g., top object at right) –Children then must choose another one that will “be the same kind of thing to eat” or that will be “the same kind of thing to play with”. –In the first case they tend to choose the object with the same color. –In the second case they will tend to choose the object with the same shape.

Adjustments to Training Environment Among the plants: –All trees are large –All flowers are small –Either can be bright or dull Among the animals: –All birds are bright –All fish are dull –Either can be small or large In other words: –Size covaries with properties that differentiate different types of plants –Brightness covaries with properties that differentiate different types of animals

Testing Feature Importance After partial learning, model is shown eight test objects: –Four “Animals”: All have skin All combinations of bright/dull and large/small –Four “Plants”: All have roots All combinations of bright/dull and large/small Representations are generated by using back-propagation to representation. Representations are then compared to see which animals are treated as most similar, and which plants are treated as most similar.

Similarities of Obtained Representations Size is relevant for Plants Brightness is relevant for Animals