CEU Electronic Theses and Dissertations, 2022
| Author | Lengyel, Gábor |
|---|---|
| Title | A Common Probabilistic Framework Explaining Learning and Generalization in Perceptual and Statistical Learning |
| Summary | Sensory learning, the process of refining perception to improve interactions with the environment in a lasting manner, is traditionally divided into two learning types: perceptual and statistical learning. The two forms of learning have been treated separately in the literature in terms of paradigms, computational models, and neural correlates. However, recent experiments eliminated the strict distinctions between PL and SL paradigms by using more complex stimuli and tasks and as a result, they found overlapping computational and neural mechanisms between the two learning types. In the present thesis, I propose a common probabilistic framework that unifies the two forms of learning and can parsimoniously explain both classical findings seemingly supporting the separation between PL and SL and more recent results demonstrating strong interactions between the two forms of learning. I argue that Hierarchical Bayesian Modeling (HBM) that inherently combines sensory bottom-up and experience-based top-down processes in a normative manner can provide a suitable unifying framework for PL and SL. In particular, HBM provides higher flexibility for efficient generalization in situations with more complex, naturalistic tasks and stimuli, displaying a hallmark of human learning. Inspired by the HBM framework, I conducted three empirical studies investigating learning and generalization in PL and SL paradigms and, after developing a computational model, I performed a simulation-based study exploring the interaction between PL and SL. In the first study, I investigated the relationship between initial performance, the amount of learning, and the extent of generalization in classical PL paradigms. I showed that (1) the previously found Weber-like relationship between initial performance and learning only shows properties of perception and does not reflect any characteristics of PL and (2) the extent of generalization was proportional to the amount of learning. Studies 2 \& 3 targeted SL focusing on how learning regularities in the stimuli influences perceptual organization. By combining the classical SL with the classical object-based attention paradigms, I showed that statistically defined chunks learnt during SL elicit object-based perceptual \& attentional effects similar to what real objects do. Next, using visual and haptic stimulation in two SL experiments, we found that participants generalized the statistically defined chunks learnt in one modality to the other modality without any learning in the other modality. These findings suggest that, relying on statistical properties, participants automatically build abstract and amodal representations of chunks that influence the segmentation of the sensory input into perceptual units. Finally, I used computational modelling to study the interaction between PL \& SL in roving paradigms and developed a unifying Bayesian Statistical Perceptual Learning model that can capture behavior in both classical and roving PL experiments. In the model, the context of the trials are inferred and the temporal transition model between the contexts is gradually learned via SL, which in turn supports the PL process modeled as an efficient resource allocation for encoding the stimuli. This interaction between PL \& SL in the model could replicate the wide range of results found in roving paradigms. Together, these results pave the road to a novel understanding of learning in vision and the concept of perceptual "objects". |
| Supervisor | Fiser, József; Lengyel, Máté |
| Department | Cognitive Science PhD |
| Full text | https://www.etd.ceu.edu/2022/lengyel_gabor.pdf |
Visit the CEU Library.
© 2007-2021, Central European University