Rekha Varrier: Never Trust the Teller! Feedback Manipulation and its Impact on Perceptual Inference
BCCN Berlin / GRK 1589 / TU Berlin
To survive and adapt in an ever-changing world, we continuously sample our surroundings through our senses. To make correct inferences about the causes of sensory signals, it is essential to learn not only about the entities present around us, but also about the reliability of sensory information itself. The influential Bayesian brain hypothesis proposes that perception is an inferential process, depending not only on sensory data, but also on beliefs about the probable causes of sensory data. Over time, the brain arrives at estimates of what to expect and how much to rely on prior beliefs and sensory data in an environment. Feedback from the environment improves learning, thereby helping the brain to arrive at these estimates. Now what will happen if the environmental feedback becomes unreliable? Providing unreliable feedback has previously been shown to impair task performance and increase pattern perception in noise. However, the mechanistic understanding about its underlying processes is currently limited. In this thesis, we explored the effects of unreliable feedback within the framework of Bayesian inference. We predicted that unreliable feedback would induce beliefs about the reliability of sensory information and lead the brain to down-weigh sensory data. To test this, we performed two behavioural experiments, a simulation and an functional magnetic resonance imaging (fMRI) experiment with visual orientation detection or discrimination tasks. We hypothesised that the sensory data would be down-weighed after an unreliable feedback phase in which invalid feedback was given in half of the trials. As a result, we predicted that task performance would deteriorate, that in the presence of prior beliefs (induced by predictive cues), perceptual inference would shift towards the priors, leading to a higher number of cue-congruent responses, and that sensory data representations in the primary visual cortex (V1) deteriorate. In one of the behavioural experiments, we additionally measured the changes in metacognitive awareness about performance, measured as confidence ratings. Reliable, correct feedback was used as a control condition in all the experiments. Data from the behavioural experiments showed that performance did indeed decrease and prior reliance increased following unreliable feedback compared to reliable feedback. Further, metacognitive awareness about performance decreased during the delivery of unreliable feedback. And lastly, the fMRI study showed that at the neural level, low-level stimulus representations deteriorated in V1 with unreliable feedback. To sum up, our results show that inducing beliefs about the reliability of sensory information by manipulating performance feedback can systematically influence perceptual inference and that these changes manifest at the earliest stages of cortical sensory processing.
PhD defense in the research training group GRK 1589, 'Sensory Computation in Neural Systems'.
Philipp Sterzer / Margret Franke
Location: BCCN Berlin, lecture hall, Philippstr. 13 Haus 6, 10115 Berlin