Fernandes, Tânia Patrícia GregórioCorreia, Luís Miguel Parreira eFottner, Nicola Alessandro2024-04-152024-04-1520232022http://hdl.handle.net/10451/64280Tese de Mestrado, Ciência Cognitiva, 2023, Universidade de Lisboa, Faculdade de CiênciasThe Neuronal Recycling Hypothesis (Dehaene, 2005; Dehaene & Cohen, 2007) proposes that the efficient computation and representation of written words at the orthographic stage of processing is enabled through the adaptation of pre-existing visual functions, which in turn, lead to the emergence of a specialised reading system. The present thesis aimed to investigate the emergence of neural detectors tuned to letters through biologically plausible computational models. A Deep Belief Network (DBN) was implemented as a model of visual shape perception, inspired by Testolin et al. (2017), and used to answer two questions: 1) does the DBN model generalise shape information that was learned from images of geometrical shapes towards classification of letters and pseudoletters (i.e., nonletters sharing the same features as letters); for example, classifying A as a triangle?; 2) is visual shape processing by a DBN sensitive to the same integration processes as those reflected in crowding effects (i.e., integration of adjacent information) by human observers; namely, by the congruency effect (better performance for targets surrounding by congruent than incongruent shapes)? The results showed that classification of letters and pseudoletters by our DBN was nonuniform across the different tested letter fonts, thus suggesting that decisions were not led by global shape. Interestingly, our model exhibited a congruence effect, and hence, a perceptual strategy similar to that previously found in illiterate adults (Fernandes et al., 2014). These results and further analyses also showed that our model’s perceptual strategy was not driven by low-level pixel similarities. The present work sets the stage to further emulate the transition from the illiterate to the ex-illiterate state, as done in the work of Hannagan et al. (2021) but with biologically more plausible learning algorithms (Bengio et al., 2015; Hinton & Salakhutdinov, 2006).engReconhecimento visual de palavrasDetector de letrasRedes neuraisAprendizagem não supervisionadaRedes Deep BeliefTeses de mestrado - 2023master thesis203524330