Integrating evidence under uncertainty: from sensory perception to reward-guided predictions

Adaptive behavior requires organisms to infer the state of the environment from information that is incomplete, noisy, and distributed across multiple sources and timescales. Although perception, learning, and decision-making are often studied as separate cognitive functions, they can also be viewed as manifestations of a common inferential problem: constructing and continuously updating internal models of the world to guide behavior under uncertainty. This thesis investigates how humans and rodents solve this problem in three complementary contexts. The first study examines temporal perception in tactile categorization tasks. Using behavioral experiments in rats and humans together with optogenetic manipulations of primary somatosensory cortex (vS1), we show that duration perception does not rely on a single fixed computational mechanism. Continuous tactile stimuli produce the well-established interaction between perceived intensity and perceived duration, consistent with a model in which duration is derived from the accumulation of sensory cortical activity. However, introducing salient onset and offset markers abolishes this interaction and renders duration judgments largely insensitive to manipulations of sensory cortical activity. These findings reveal the existence of at least two distinct strategies for temporal estimation whose recruitment depends on stimulus structure and sensory context. The second study investigates adaptive inference from reward-history information. Rats and humans performed a two-alternative foraging task in which reward locations followed probabilistic sequential patterns generated by a hidden Markov process. Both species successfully learned and exploited these regularities and adapted their behavior when environmental statistics changed. Across species, behavior exhibited a robust asymmetry between responses following rewarded and unrewarded outcomes, with rewarded outcomes exerting a stronger influence on subsequent choices. The third study examines how sensory evidence and reward-history information are combined during decision-making. Human participants performed a task in which tactile cues and probabilistic reward patterns jointly predicted reward location. Performance exceeded that observed when either information source was available alone, indicating that individuals integrate sensory and reward-based information rather than relying on either source independently. Psychophysical analyses further suggest that choices are guided by a unified decision variable combining current sensory evidence with beliefs derived from recent experience. Together, these studies demonstrate that adaptive behavior emerges from flexible mechanisms that integrate information across sensory signals, reward histories, and contextual factors. These findings support the view that perception, learning, and decision-making are unified by common principles of adaptive inference under uncertainty.