Comparing Early Visual Representations of Letters in Dyslexic and Typical Readers

16:30 - 18:00

Symbolic Representations in the Visual Pathway: From Numbers and Letters to Words and Semantics

Room: HSZ - N3

Chair/s:

Jack E Taylor, Janos Pauli

Written script, including numbers, letters, and letter-strings, is a recent cultural invention central to letter- and number-literate societies. In such societies, humans learn early on to recognise glyphs and map them fluently onto specific sounds and concepts. This symposium explores how the brain achieves this objective, using multiple complementary lenses to understand the processing of linguistic and mathematical symbols, the degree to which these representations are distinct, and interactions between visual symbol recognition and abstract processes of language and numerosity.

The first talk introduces a predictive coding-motivated computational model of letter recognition, showing how these principles might explain the recognition of letters in noisy environments. This work suggests that predictive coding accounts of word recognition may also apply to isolated letters. The second talk uses an optimal transport framework to model the space of early visual representations of letter symbols revealed by EEG, exploring how such representations may be altered in dyslexia. This work tests whether this learning disorder in reading also results in weaker neural alignment with computational models of letter representations. The third talk presents an analysis of human fMRI and macaque electrocorticography responses to naturalistic images, suggesting a shared prominent representation of stimuli related to both orthography and numerosity. This finding is discussed in relation to the notion of proto-architecture for mathematical cognition in the higher-level visual cortex of non-human primates. The fourth talk examines interactions between the processing of Arabic digits and language. This study exploits the discrepancy between the base-10 system of Arabic numbers and base-20 system of French number words, finding that native French speakers utilise language during a numerical task, even when language is redundant. The fifth talk explores how the brain processes words with varying degrees of misspelling. Using MEG data, it examines connectivity between lower visual areas and the lvOT, suggesting that lvOT processes real words in a feedforward manner but engages feedback mechanisms for misspelled words and pseudowords.

Combining experimental and computational approaches, this symposium advances our understanding of how the brain maps arbitrary visual forms into meaningful symbolic representations, and how these processes interact with language and numerosity.

Submission 507

Comparing Early Visual Representations of Letters in Dyslexic and Typical Readers

SymposiumTalk-02

Presented by: Jack E. Taylor

Jack E. Taylor ^{1, 2, 3}, Caroline Bergmann ¹, Christian J. Fiebach ^{1, 2}

¹ Department for Psychology, Goethe University Frankfurt, Germany

² COBIC Brain Imaging Center, Goethe University Frankfurt, Germany

³ School of Psychology and Neuroscience, University of Glasgow, United Kingdom

The earliest stages of visual word recognition involve recognising the shapes of letters and characters that make up words. In this preregistered study, we examine whether such early neural representations of letter shapes differ between 20 participants with Dyslexia and 20 controls, matched on age, gender, education, and non-verbal IQ. In a Bayesian, hierarchical representational similarity analysis of electroencephalography (EEG) data, we examine group differences in brain-model alignment. We focus on a recent computational framework built on optimal transport theory, in which representational similarity is proportional to a minimum edit distance between two-dimensional histograms of letters. A key feature of this framework is that it can be used to capture a transition from more feature-based, retinotopic representations, to more structural representations, invariant to transformations like translation, rotation, scaling, and mirroring. We apply this approach to examine (1) whether the typical-reading control participants replicate previous results, (2) whether the computational models align better or more poorly with the neural activity of dyslexic participants, and (3) whether dyslexic participants differ in the degree of invariance to transformations. Data collection is ongoing, but current results from dyslexic participants tested so far suggest that their neural representations of letters align less well than expected (based on the prior from the previous study) with the optimal transport models, including reduced evidence for invariant representations. These initial findings tentatively suggest that even early neural representations of single letters may differ in dyslexia.

Bookmark