Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

A Reinforcement Learning Model for Task-modulated Perception

Mohammadreza Bigham¹, Shervin Safavi²; ¹Shahid Beheshti University, ²Technische Universität Dresden

Presenter: Mohammadreza Bigham

Perception is a highly active process; therefore, it can and should be approached from a decision-theoretic perspective. Elevating perception from a mere sensory inference to a decision process requires us to consider, for instance, how the value of sensory objects influences what we perceive and how the task at hand affects our perception. Here, we suggest that multistable perception could be a suitable candidate to study task-modulated perception in both humans and animals. Multistable perception is the dynamical alternation that arises when a single sensory input has more than one interpretation or explanation. Multistable perception is one of the most venerable perceptual phenomena that has been formalized as a decision process. We extend the previous model of perceptual multistability by incorporating a richer state space and action repertoire for a reinforcement learning agent, and we show that this allows us to explain the established task-modulated perception during perceptual multistability. Our model replicates and explains recent findings on the modulation of perception by task observed in previous studies. This is achieved by incorporating two key elements-- changes in attentional resource allocation and representation of the environment volatility-- into a reinforcement learning paradigm. These changes are implemented in the model by systematically adjusting the observation and transition functions in our partially observable Markov decision process (POMDP) model of perpetual decisions. Taken together, our findings further support the view that perception is an active, goal-directed process, aligned with principles shared by other aspects of cognition.

Topic Area: Reward, Value & Social Decision Making

Extended Abstract: Full Text PDF