Neuro-X Seminar: Rational Sensing from insects to rodents to humans to machines
Event details
| Date | 02.03.2023 |
| Hour | 13:00 › 14:00 |
| Speaker | Prof Rafael Polania |
| Location | Online |
| Category | Conferences - Seminars |
| Event Language | English |
NOTE: for logistical reasons, please do access the seminar room by the exterior of AI extension (level 0, entrance AI 0354)
Is the role of our sensory systems to represent the physical world as accurately as possible? If so, are our preferences and actions—which are often labeled as irrational—decoupled from these “ground-truth” sensory experiences? We argue that the answer to both questions is no. Perhaps counterintuitively, we propose that accurate representations of sensory signals do not necessarily maximize the organism’s chances of survival. To test this hypothesis, we developed a unified normative framework for fitness-maximizing encoding by combining theoretical insights from neuroscience, computer science, and economics. Initially, we applied predictions of this model to neural responses from large monopolar cells (LMCs) in the blowfly retina. We found that neural codes that maximize reward expectation—and not accurate sensory representations—account for retinal LMC activity. Behavioral experiments in humans revealed that sensory encoding strategies are flexibly adapted to promote fitness maximization. Moreover, human fMRI data confirmed that novel behavioral goals that rely on object perception induce efficient stimulus representations in early sensory structures. Interestingly, this result was confirmed by deep neural networks with information capacity constraints trained to solve the same task in humans. Furthermore, experiments in which rodents were trained to solve the same task in humans revealed that mice also adaptively allocate their sensory resources in a way that maximizes reward consumption in novel stimulus-reward association environments. These experiments allowed us to discover that arousal systems carry reward distribution information of sensory signals and that distributional reinforcement learning mechanisms—a fundamental mechanism in state-of-the-art machine learning algorithms—regulate sensory precision via top-down normalization. These findings reveal how agents can efficiently perceive and adapt to environmental contexts within the constraints imposed by neurobiology. Thus, the often-observed irrationalities and biases attributed to downstream processing might unavoidably originate from the way early sensory systems should adapt to and process information in insects, rodents, humans, and machines.
Is the role of our sensory systems to represent the physical world as accurately as possible? If so, are our preferences and actions—which are often labeled as irrational—decoupled from these “ground-truth” sensory experiences? We argue that the answer to both questions is no. Perhaps counterintuitively, we propose that accurate representations of sensory signals do not necessarily maximize the organism’s chances of survival. To test this hypothesis, we developed a unified normative framework for fitness-maximizing encoding by combining theoretical insights from neuroscience, computer science, and economics. Initially, we applied predictions of this model to neural responses from large monopolar cells (LMCs) in the blowfly retina. We found that neural codes that maximize reward expectation—and not accurate sensory representations—account for retinal LMC activity. Behavioral experiments in humans revealed that sensory encoding strategies are flexibly adapted to promote fitness maximization. Moreover, human fMRI data confirmed that novel behavioral goals that rely on object perception induce efficient stimulus representations in early sensory structures. Interestingly, this result was confirmed by deep neural networks with information capacity constraints trained to solve the same task in humans. Furthermore, experiments in which rodents were trained to solve the same task in humans revealed that mice also adaptively allocate their sensory resources in a way that maximizes reward consumption in novel stimulus-reward association environments. These experiments allowed us to discover that arousal systems carry reward distribution information of sensory signals and that distributional reinforcement learning mechanisms—a fundamental mechanism in state-of-the-art machine learning algorithms—regulate sensory precision via top-down normalization. These findings reveal how agents can efficiently perceive and adapt to environmental contexts within the constraints imposed by neurobiology. Thus, the often-observed irrationalities and biases attributed to downstream processing might unavoidably originate from the way early sensory systems should adapt to and process information in insects, rodents, humans, and machines.
Practical information
- Informed public
- Free