2:00pm, Wednesday, 30 January 2008:
<br>
<br>
<center><b>
Bayesian perception
</b>
<br>
<br> 
Alan Stocker
<br>
<br>
</center>

Generating a sensible and stable percept of the world is
crucial. Ambiguities, as well as noise and other sensory limitations
make this a hard computational problem for the human brain. Yet
evolution presses for optimal solutions, giving rise to the hypothesis
that perception is the process of optimal statistical inference
(combining noisy sensory evidence with prior assumptions about the
world). Based on this hypothesis, I will formulate a Bayesian observer
model for human visual motion perception. The model not only accounts
for the average bias and trial-to-trial variability in subjects'
perceived speed, but it also allows us to reverse-engineer the exact
form of the subjects' prior assumptions and noise characteristics from
the perceptual data. Such quantitative characterization is critical
for validating the Bayesian hypothesis. I will also present recent results
of cross-validating the extracted prior and noise characteristic by
comparing model predictions with subjects behavior in a completely
different psychophysical motion experiment. Finally, I will address
some limitations of the Bayesian modeling approach that are revealed
in recently reported psychophysical experiments. I will show that
human subjects exhibit a strong tendency to abandon the optimal
Bayesian solution in order to maintain a consistent assessment of the
sensory evidence. Interestingly, this behavior parallels human
avoidance of cognitive dissonance, suggesting functional similarities
between low-level perception and cognition.