Modeling dyslexia using vision language models

Event details
Date | 18.08.2025 |
Hour | 13:00 › 15:00 |
Speaker | Melika Honarmand |
Location | |
Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Tanja Käser
Thesis advisor: Prof. Martin Schrimpf
Co-examiner: Prof. Marcel Salathé
Abstract
Dyslexia, a cognitive disorder affecting up to 15% of individuals worldwide, impairs reading and writing abilities but has no effect on non-verbal measures of intelligence. While traditionally considered a phonological disorder, recent evidence highlights deficits in visual processing, particularly within the visual word form area (VWFA). To elucidate the mechanistic underpinnings of dyslexia, we here employ vision-language models (VLMs) as computational analogs to human neural and behavioral responses. We evaluate VLM performance on two benchmarks corresponding to reading and non-verbal performance measures: the Rapid Online Assessment of Reading (ROAR), a visual test distinguishing real words from pseudowords, and RAVEN Progressive Matrices, a measure of non-verbal IQ. Our models demonstrate high accuracy on both tasks, consistent with the performance of healthy subjects. We generate candidate dyslexia models by identifying VWFA-analogous units in the models and muting their activity. On the one hand, these models experience a significant decline in ROAR performance to dyslexic levels while preserving RAVEN scores. On the other hand, lesioning a comparable number of random units barely affects ROAR and RAVEN performance. By building end-to-end computational models of dyslexia, our findings provide mechanistic insights into the causal role of VWFA impairments. Beyond a computational understanding of the neural underpinnings of dyslexia, these models lay the foundation for novel potential treatment and screening strategies.
Selected papers
coming soon
Exam president: Prof. Tanja Käser
Thesis advisor: Prof. Martin Schrimpf
Co-examiner: Prof. Marcel Salathé
Abstract
Dyslexia, a cognitive disorder affecting up to 15% of individuals worldwide, impairs reading and writing abilities but has no effect on non-verbal measures of intelligence. While traditionally considered a phonological disorder, recent evidence highlights deficits in visual processing, particularly within the visual word form area (VWFA). To elucidate the mechanistic underpinnings of dyslexia, we here employ vision-language models (VLMs) as computational analogs to human neural and behavioral responses. We evaluate VLM performance on two benchmarks corresponding to reading and non-verbal performance measures: the Rapid Online Assessment of Reading (ROAR), a visual test distinguishing real words from pseudowords, and RAVEN Progressive Matrices, a measure of non-verbal IQ. Our models demonstrate high accuracy on both tasks, consistent with the performance of healthy subjects. We generate candidate dyslexia models by identifying VWFA-analogous units in the models and muting their activity. On the one hand, these models experience a significant decline in ROAR performance to dyslexic levels while preserving RAVEN scores. On the other hand, lesioning a comparable number of random units barely affects ROAR and RAVEN performance. By building end-to-end computational models of dyslexia, our findings provide mechanistic insights into the causal role of VWFA impairments. Beyond a computational understanding of the neural underpinnings of dyslexia, these models lay the foundation for novel potential treatment and screening strategies.
Selected papers
coming soon
Practical information
- General public
- Free
Contact
- edic@epfl.ch