Ideal Bayesian learning in human scene perception
Lists
Only you see this
Abstract
Humans are confronted with massive amounts of visual information that must be collapsed into meaningful chunks, describing which elements of a visual scene belong together, for efficient recognition of faces or objects. This entails the classic model-selection trade-off: choosing the right inventory of chunks that accurately captures the structure of scenes already observed but is not too specific to their details so that it also generalizes well to new scenes. Here we address this problem by asking two questions: what are the optimal learning strategies for the formation of visual chunks, and do humans employ these strategies during unsupervised learning? We implement an ideal learner that uses Bayesian model comparison, a principled statistical solution for the model-selection problem, to extract and store only those chunks that are minimally sufficient to encode the scenes. We then contrast this model with previous pair-wise associative and iterative accounts of learning in a psychophysical experiment. We show that human performance is well above chance when pair-wise statistics in the scenes contain no relevant information. In contrast to previous accounts of human learning, our ideal Bayesian learner can reproduce the result of the experiment, as well as a large set of previous empirical findings. These results suggest that humans learn new complex information from visual scenes by generating economical representations based on previous experience and not by encoding the full correlational structure of the input.
BibTeX
@inproceedings{fiser2007ideal,
author = {Jozesef Fiser and Gergő Orbán and RN Aslin and Máté Lengyel},
title = {{Ideal Bayesian learning in human scene perception}},
booktitle = {Cambridge},
year = {2007},
}Loading related content...