Optimal computation of flow field variables from natural visual signals.
R.R. de Ruyter van Steveninck, W. Bialek
Many animals use visual information to navigate their environment, and in this context it is important for them to estimate how they move through space. Visual input, gathered by the eye, contains information that is related to self motion, but the connection between what we see and how we are moving is indirect and sometimes ambiguous. Moreover, light is carried by photons, and because these arrive at random the visual input forms a noisy representation of the surroundings. To make optimal estimates of self motion from visual input, the brain must therefore use an algorithm that takes into account the statistics of the visual input signals and the probabilistic relation between visual input and self motion. This project will study motion estimation in the natural world as a statistical estimation problem, and compare the predictions of statistically optimal processing to measurements in a biological system. There are fundamental as well as practical reasons for studying the problem, but there is an additional motivation. That is to investigate if, and to what extent, neural computations in biological systems can be understood as being optimized for their specific tasks in the natural environment. This is a hard issue to solve in general, since the answer depends on poorly known statistical properties of natural sensory signals. The present case provides an example where one can measure and quantify both the signals that need to be estimated (rotations) and the data on which this estimate is based (raw visual input). Moreover, a level of statistical sampling can be achieved that allows a direct application of statistical inference to data that are representative for natural sensory signals. The investigators will construct a precise high speed camera, with spatial sampling characteristics representative for the fly visual system, and with associated gyrosensors to measure camera rotation along three axes (yaw, pitch and roll). With this camera they will make precise simultaneous measurements of rotational camera motion and visual input in a natural environment. This simultaneous sampling makes it possible to effectively measure the probability distributions that describe the relation between motion and visual input. From this distribution, the characteristics of the optimal statistical motion estimator can be derived. These predictions for the computational structure of the optimal estimator will be compared to the behavior of motion sensitive neurons recorded from the visual brain of the blowfly. That comparison will allow quantitative assessement of the extent to which biological motion processing in the blowfly approaches optimal performance. Preliminary experiments have shown that both the blowfly and the optimal estimator show specific biases in their output, depending on the statistics of the input signals. There are strong indications that other animals, including humans, share similar biases, suggesting that these biases are an inevitable and universal consequence of the optimal processing of natural sensory signals.
