Probabilistic inference in the primary visual cortex

• Applicant: Professor Dr. Alexander Ecker
• Subject Area: Cognitive, Systems and Behavioural Neurobiology
• Term: from 2015 to 2020
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 270450705

Nearly 150 years ago, Hermann von Helmholtz conjectured that visual perception is the result of an inference process, thus challenging the traditional view of perception as a passive window to the outside world (Helmholtz, 1867). He postulated that our brain has learned an internal, generative model of the world that we use to perform perceptual inference. The present proposal builds upon this idea and aims at understanding the neural mechanisms underlying this inference process. We employ the mathematical framework of probabilistic (Bayesian) inference to study how the brain implements visual perception and decision-making. We hypothesize that each brain area computes a posterior over some set of features it represents by combining sensory evidence (bottom-up, likelihood) with internal beliefs and knowledge about the state of the world (top-down, prior). Since the subject's internal belief, which includes the focus of attention and expectations, is a highly dynamic process that is never identical on different trials, this hypothesis predicts that neuronal responses should vary from trial to trial and this variability should contain behaviorally relevant information. In the present proposal, we use multi-electrode recordings in V1 of behaving monkeys and mathematical state space models to infer the subject's internal belief from the trial-to-trial variability in neural population responses. We conceptualize spatial attention as a prior on location and use a change detection paradigm that induces varying degrees of fluctuation in the attentional signal by changing whether subjects must attend to one location while ignoring another, or attempt to attend to both locations simultaneously. Since fluctuations in the attentional signal should increase as the subjects need to split their attention, correlated neuronal variability should increase proportionally. In Objective 1 we use Gaussian Process Factor Analysis to infer the subject's attentional state in real-time, allowing us to characterize its spatiotemporal dynamics both at the neuronal and the behavioral level. In Objectives 2 and 3 we test two key predictions of the probabilistic inference framework: (a) the relative weighting of the internal belief as we vary the strength of the sensory evidence, and (b) the temporal dynamics of the attention signal for different stochastic stimulus sequences. Ultimately, we hope that the probabilistic inference framework can serve as a normative account for a variety of cognitive processes and that our experimental and theoretical approach will allow us to observe their temporal dynamics in real-time and read the subject's state of mind on a single-trial basis. Such knowledge eventually promises to help us develop better diagnostic markers for a variety of brain disorders and to develop a mechanistic understanding of how visual and cognitive processing is disturbed by such disorders at the network level.

• DFG Programme: Research Grants