A central component of the predictive coding theoretical framework concerns the comparison between predictions and sensory decoding. In the probabilistic setting, this takes the form of assessing the similarity or distance between probability distributions. However, such similarity or distance measures are not associated with explicit probabilistic models, making their assumptions implicit. In this paper, we explore an original variation on probabilistic coherence variables; we define a probabilistic component, that we call a "Bayesian comparator", that mathematically yields a particular similarity measure. A geometrical analogy suggests two variants of this measure. We apply these similarity measures to simulate the comparison of known, predicted patterns to patterns from sensory decoding, first in a simple, illustrative model, and second, in a previous model of visual word recognition. Experimental results suggest that the variant that is scaled by the norms of both predicted and perceived probability distributions yields better robustness and more desirable dynamics.

To explain how perception processes are performed, understanding how continuous sensory streams are temporally segmented into discrete units is central. This is particularly the case in speech perception where temporal segmentation is key for identifying linguistic units contained between consecutive events in time. We propose an original probabilistic construct, that we call "Bayesian gates", to segment temporally continuous streams of sensory stimuli into sequences of decoders. We first define Bayesian gates mathematically and describe their properties. We then illustrate their behavior in the context of a model of word recognition in speech perception. We show that, based on an event detection module, they sequentially parse the acoustic stimulus, so that each syllable decoder only processes a segment of the sensory signal.

A new Bayesian model of visual word recognition is used tosimulate neighborhood frequency effects in lexical decision.These effects have been reported as being either facilitatory orinhibitory in behavioral experiments. Our model manages tosimulate the apparently contradictory findings. Indeed, study-ing the dynamic time course of information accumulation inthe model shows that effects are facilitatory early, and becomeinhibitory at later stages. The model provides new insights onthe mechanisms at play and their dynamics, leading to betterunderstand the experimental conditions that should yield a fa-cilitatory or an inhibitory neighborhood frequency effect.

In visual word processing modeling, few models have success-fully accounted for a large variety of tasks, and large corpora ofbehavioral observations. We consider a dataset from a megas-tudy, in which participants performed three tasks (lexical de-cision, word naming, and word recognition in a progressivedemasking situation), on the same, large set of stimuli. Wedefine the BRAID-Phon model, an extension of a previousprobabilistic model, the BRAID model, whose originality isits visuo-attentional component, in which a visuo-attentionaldistribution spatially deploys sensory processing capabilities.BRAID-Phon includes phonological representations of words,allowing simulating the naming task. We simulated the threetasks on the dataset we considered, and analyzed predicted re-action times in terms of word frequency and word length ef-fects. Simulation results show that BRAID-Phon successfullycaptures the direction and order of magnitude of the observedeffects, in all three tasks.

Reading acquisition primarily relies on orthographic learning. Behavioral studies show that familiarity with a novel word's pronunciation facilitates learning, particularly in semantically meaningful contexts. Two main components of orthographic learning are commonly described: perceptual processing of the visual stimulus, to infer corresponding phonological rep- resentations, and “pronunciation correction”, to correct errors from perceptual processing. Currently, pronunciation correc- tion has not been featured in reading acquisition computa- tional models. This study uses BRAID-Acq, a reading ac- quisition model, to implement and compare two pronunciation correction mechanisms (an “online” and a “post-processing” variant). We simulated learning of words with and without prior phonological knowledge and explored the impact of con- text strength and size on learning. Results indicate that both mechanisms improve decoding. However, the post-processing mechanism induced implausible lexicalization for words with- out prior phonological knowledge, while the online mecha- nism did not. Overall, our simulation results suggest that pro- nunciation correction could be construed as an online process.

This study focuses on inter-individual convergence effects in the perception and categorization of speech sounds. We ask to what extent two listeners can come to establish a shared set of categorization criteria in a phoneme identification task that they accomplish together. Several hypotheses  are  laid  out  in  the  framework  of  a  Bayesian  model  of  speech  perception  that  we  have  developed  to  account  for  how  two  listeners  may  each  infer  the  parameters  that  govern  their partner’s responses. In our experimental paradigm, participants were asked to perform a joint  phoneme  identification  task  with  a  partner  that,  unbeknownst  to  them,  was  an  artificial  agent, whose responses we manipulated along two dimensions, the location of the categorical boundary and the slope of the identification function. Convergence was found to arise for bias but not for slope. Numerical simulations suggested that lack of convergence in slope may stem from the listeners’ prior level of confidence in the variance in VOT for the two phonemic categories. This study sheds new light on perceptual convergence between listeners in the categorization of speech sounds, a phenomenon that has received little attention so far in spite of its central importance for speech communication.