Prof. Dr. Felix Wichmann
Field of Research
My group's research goal is to understand visual perception. We investigate spatial vision from the level of simple artificial stimuli up to object and scene perception, combining psychophysical experiments and computational modelling. Currently, we have three research foci:
First, we attempt to uncover the critical features that observers use in complex perceptual tasks. Given the high-dimensional visual input, what are the features on which the visual system bases its computations? The lack of knowledge about the critical features constitutes one of the major obstacles preventing the development of successful computational models of visual perception. In my group we develop inverse machine learning methods for nonlinear systems identification to find the critical features.
Second, we work on a computationally efficient implementation of a psychophysical model of early spatial vision. Behavioural pattern detection experiments have greatly advanced our understanding of the computations performed by the early visual system to extract information about the environment from the retinal image. Up to now, psychophysical near-threshold measurements have been taken to suggest that observers select the maximum response from a bank of parallel linear visual filters (each sensitive to a specific image resolution) to perform detection. However, recent insights into the neurophysiological underpinnings of vision are hard to reconcile with this view. We have shown that maximum-output decoding of linear and independent spatial frequency channels falls short of a complete explanation of pattern detection when challenged to fit a broad range of classic detection phenomena using a single set of parameters. Maximum-likelihood read-out of a neurophysiologically-inspired model of population activity in the primary visual cortex, on the other hand, can fully account for pattern detectability as investigated in behavioural detection, summation, adaptation and uncertainty experiments. This is important in its own right, but in addition such a model should provide us with realistic constraints for the machine learning methods developed for feature identification, and help us to incorporate domain knowledge into the pre-processing stage of our system identification algorithms.
A related issue is assessing to what degree insights obtained using simple, well-controlled stimuli used in early spatial vision research generalize to natural, more complex settings. Here we have recently shown that so-called “natural images” may be less natural than commonly presumed.
Third, we are interested in improving psychophysical methods to obtain more reliable data for our models. In behavioural studies sensitivity to a certain stimulus property is typically measured over a period of time. When analysing the data most studies assume stationary observers: responses are expected to be dependent only on the intensity of a presented stimulus and not on other factors such as stimulus sequence, duration of the experiment, or the responses on previous trials. Unfortunately, a number of factors such as learning, fatigue, or fluctuations in attention and motivation will typically result in violations of this assumption. Recently we developed a simple adjustment of the confidence intervals that corrects for the estimation error almost independently of the number of trials and the particular type of violation.