Computational simulation of reading acquisition and of dyslexia

In order to test precise hypotheses about learning mechanisms and the causes of cognitive disorders, it is increasingly useful to have computational models of the cognitive functions concerned.
In the field of reading, models traditionally fall into two main categories: historical connectionist models which make a minimum of assumptions and implement a minimum of representations, but which, trying to explain too much with too little, produce fairly crude results which fail to model the most subtle phenomena of reading and its acquisition. On the other hand, there are more sophisticated, symbolic or connectionist models, which aim to simulate a wider range of phenomena, but at the cost of ad hoc specifications which appear unrealistic. The obvious limitations of these two categories of models offer little prospect of simulating learning situations, cognitive tasks and reading deficits in a sufficiently accurate and realistic way to be relevant.
The BRAID family of reading models, developed by Julien Diard and Sylviane Valdois with 4 successive PhD students, offers an architecture that overcomes these limitations and makes it possible to simulate learning to read and reading difficulties in all their complexity.
In particular, the Bayesian BRAID-Acq model recently developed by Steinhilber, Valdois and Diard can be used to simulate the second phase of learning to read, in which the child, already knowing a certain number of written words (>1000), is able to read independently and gradually increase her orthographic lexicon as she is exposed to new words (unsupervised learning).

The first objective will be to adapt the model to simulate the 1st phase of learning, during which the child learns the grapheme-phoneme correspondences to decode words, gradually storing the decoded words in her orthographic lexicon. The model can then be used to compare different methods of teaching reading: different progressions of grapheme/phoneme correspondences and words used in different textbooks, analytical vs. synthetic methods, global methods. It could also be used to test the value of certain specific ingredients, such as the importance of supervised (rather than incidental) learning of grapheme-phoneme correspondences, the positive or negative effect of words learned globally in pre-school before learning to read, or the global learning of frequent whole words. The model's predictions can be compared with existing behavioural data or data to be collected from beginning readers.
Once the process of learning to read has been realistically modelled from beginning to end, a second objective will be to simulate different hypotheses about the cognitive causes of dyslexia: to simulate different types of phonological, visual or visual-attentional deficits, which have already been postulated, and to study their consequences on learning to read and on performance in different reading tasks. Once again, these predictions can be compared with reading data and other tasks obtained from dyslexic pupils. Ultimately, the aim will be to generate new predictions that can be tested by new behavioural experiments, and thus to disentangle the different cognitive theories of dyslexia. It will also be possible to test the hypothesis that the population of dyslexic children is made up of several sub-populations with different cognitive deficits, in proportions to be determined.

Candidate profile:
The candidate should ideally have a strong background in computer programming and simulation, and a taste for mathematical modelling and experimental psychology: students in computer science, engineering and cognitive science. Knowledge in probability would be an asset. The model is developed in Python, so familiarity with this language is a prerequisite.

Contacts
Franck Ramus (LSCP, CNRS) : franck.ramus@ens.psl.eu
Julien Diard (LPNC, CNRS) : julien.diard@univ-grenoble-alpes.fr