Computational models of reading.
Probablistic Graphical Models of Dyslexia: Reading is a complex cognitive process, errors in which may assume diverse forms. To capture the complex structure of reading errors, we propose a novel way of analyzing reading errors made by dyslexic people, using probabilistic graphical models. Our study focuses on three inquiries. (a) We examine which graphical model best captures the hidden structure of reading errors. (b) We examine whether a graphical model can diagnose dyslexia closely to how experts do (c) We draw on our results to shed light on a theoretical debate about the definition and heterogeneity of dyslexia. To address these questions, we explore three different models, an LDA-based model and two Naive Bayes models which differ by their assumptions about the generation process of reading errors. The models are trained on a large corpus of reading errors. Our results show that a Naive Bayes model best agrees with labels given by clinicians and can be therefore used for automation of the diagnosis process. The LDA-based model best captures patterns of reading errors and could therefore contribute to the understanding of dyslexia and to the diagnostic procedure. Finally, our results support a model assuming multiple dyslexia subtypes, that of a heterogeneous view of dyslexia.
Metric learning for phoneme representations: Theories of phoneme representations have proposed that subphonemic features underlie perceptual differences between phonemes. However, in coming to explain perceptual confusion between phonemes, existing theories fail to account for actual confusion rates collected from subjects. To address this gap, we adopt a new approach to better characterize the role of subphonemic features in the way people discriminate between phonemes. We explore data-driven metrics that are defined over subphonemic features and are learned from phoneme-confusion data. Weights assigned by the function to each subphonemic feature indicate its role in discriminating between phonemes. Results show that such metrics enhance prediction of unseen perceptual distances between phonemes, in comparison to theoretical metrics previously proposed in the literature. Analyses of the learned metric functions reveal an order between subphonemic features with respect to their perceptually-discriminative power. The leading subphonemic features for English are, in descending order: voicing, nasality, distributed-stridents and approximants. The learned metric function thus quantifies the contribution of subphonemic features to the overall perceptual distances. Finally, we assess the degree to which metric functions differ across languages. To do this, we collect phoneme-confusion data in Hebrew and analyze the differences between the Hebrew and English metrics. Results point to differences between languages and the need for language-specific, rather than universal, assessments.
A biologically-inspired model of visual word recognition: We present a computational model of visual word recognition. The model is biologically inspired, incorporating plausible cortical dynamics, thus adding to previous studies, which have used connectionist or 'box-and-arrow' type models. We begin by exploring several methods to represent the letter identities in an artificial neural network, and to identify the method that best agrees with experimental findings and computational constraints. In the self-organization process of a multilayer neural network, letter-identity and letter-position representations are further processed to create word representations. These correspond to word memories in an orthographic lexicon, as described in neuropsychological models, and function as attractors of the neural network. Simulations present normal reading by the network in the absence of noise or deficits. When noise or deficits are introduced, the network presents failures such as letter transposition or letter substitution, which are similar to those made by dyslexics with letter-position dyslexia and letter-identity dyslexia, respectively.
A neuronal-based model for the process of reading (ENCODS, BORDEAUX 2013).
Probabilistic graphical models of dyslexia. KDD2015, Sydney, Australia.
A biologically-inspired model of visual word recognition. Presented at the Belgrade BioInformatics Conference (BelBI2016), Belgrade, Serbia.
The perceptual structure of the phoneme manifold. Trieste Encounters on Cognitive Science (TEX2016), Trieste, Italy
Lakretz, Y., Chechik, G., Friedmann, N., Rosen-Zvi, M. (2015). Probabilistic Graphical Models of dyslexia, in submission to KDD 2015.
Lakretz, Y., Marjanovic, K., Gu, Y., Treves, A. (2015). Distinguishing between different syntactic roles of identical words in normal reading: an ERP study. CogSci, 2015.
Lakretz, Yair, Naama Friedmann, and Alessandro Treves. "A biologically-inspired model of visual word recognition." Belgrade BioInformatics Conference 2016.
Lakretz, Y., Chechik G., Cohen-Gary E., Treves A., Friedmann N. (in preparation). The perceptually discriminative power of subphonemic features.
2010 – The Hebrew University, department of Philosophy – Best Thesis award.
2012 – Tel-Aviv University, Sagol Seminars’ Best Lecture Award, accompanied by a travel prize.
2014 – Tel-Aviv University – Travel fellowship
2015 – Visiting PhD student fellowship, SISSA, Italy.
2015 - Travel award to KDD2015 conference, Sydney, Australia.
2016 - Rector’s award for excellence in teaching, Tel-Aviv University.
Research Categories: Cognitive neuroscience, Computational neuroscience