Revisiting the Role of Relearning in Semantic Dementia

TL;DR

The paper investigates whether SD-like semantic deterioration arises from atrophy alone or from a combination of atrophy and relearning. Using a two-layer linear network whose hidden layer models the anterior temporal lobe, the authors demonstrate that deleting hidden units to simulate atrophy only yields SD-inconsistent patterns, whereas incorporating retraining (relearning) after each deletion reproduces category-coordinate, cross-category, and prototyping errors observed in SD. The findings argue that continual relearning of lost information, rather than atrophy per se, can drive the progression of semantic deficits, even in the absence of output non-linearities. This proposes a broader framework for cognitive disease progression and highlights relearning as a potential target for future research across chronic conditions.

Abstract

Patients with semantic dementia (SD) present with remarkably consistent atrophy of neurons in the anterior temporal lobe and behavioural impairments, such as graded loss of category knowledge. While relearning of lost knowledge has been shown in acute brain injuries such as stroke, it has not been widely supported in chronic cognitive diseases such as SD. Previous research has shown that deep linear artificial neural networks exhibit stages of semantic learning akin to humans. Here, we use a deep linear network to test the hypothesis that relearning during disease progression rather than particular atrophy cause the specific behavioural patterns associated with SD. After training the network to generate the common semantic features of various hierarchically organised objects, neurons are successively deleted to mimic atrophy while retraining the model. The model with relearning and deleted neurons reproduced errors specific to SD, including prototyping errors and cross-category confusions. This suggests that relearning is necessary for artificial neural networks to reproduce the behavioural patterns associated with SD in the absence of \textit{output} non-linearities. Our results support a theory of SD progression that results from continuous relearning of lost information. Future research should revisit the role of relearning as a contributing factor to cognitive diseases.

Figures (2)

• Figure 1: As with a patient with SD ((b) taken from hodges1995charting), the model (a) makes category coordinate errors first and progresses to cross-category and superordinate errors. Additionally, if odd data points are shown twice as often during relearning ($\hat{Y}_{relearn\text{-}freq}$) then the more frequently seen features will dominate the representations. As a result, the less frequent objects are mistaken for the prototypical ones when making category coordinate and cross-category errors. Thus, a linear network with relearning reproduces the pattern of errors and prototyping effect associated with SD in humans.
• Figure 2: Setup and Primary Results: The linear network ($W^2W^1$) learns to map hierarchically structured data ($X$) to the corresponding features ($Y_{true}$). We model SD by deleting hidden neurons after training is complete. We compare two models: one with no relearning ($\hat{Y}_{base}$) and one with relearning ($\hat{Y}_{relearn}$). A model with no relearning loses information across all levels of the hierarchy at once, contradicting the patterns associated to SD. A model with relearning loses specific feature information before general features, consistent with what is expected of SD.