Why do we see it that way ? Testing physics-based theories of visual computation

William Bialek (Princeton University, USA City University of New York) - 15 mai 2014.

Helmholtz understood that perception is inference. What we "see" is not the raw data collected by the array of detectors in our eyes, but an interpretation. This is, perhaps, not unlike the problem of interpreting the data from a complex physics experiment. Indeed, there has been considerable interest in the idea that our brains interpret raw sense data using proper probabilistic models, combining noisy data with prior knowledge to make nearly optimal estimates of quantities that matter, such as motion or depth. These ideas can be a given a very precise formulation using the tools of statisti- cal physics. The difficulty is that the resulting theory has almost no predictive power unless we know something about the noise levels of the detectors and the a priori distributions of the relevant signals. I will focus on the example of visual motion estimation, especially in flies, where we have known for many years that the system makes nearly optimal estimates, being limited primarily by photon shot noise and diffraction through the small lenses of the compound eye. In this system we know a great deal about the noise levels in the detectors, and very recent experiments are making it possible to sample the movies that fall on the retina and the associated motion signals. Armed with these new data, theories from twenty years ago now generate precise predictions. I will highlight some surprises : the visual data we collect are surprisingly impoverished, and as a result the physically optimal estimates are systematically distorted in ways that have long been through to be symptoms of biological limitations.