Understanding how the brain transforms often uncertain sensory evidence into a perceptual decision, and how confidence in that decision emerges, remains a core challenge in cognitive neuroscience. To fully explain how external inputs, prior expectations, and internal arousal shape perception and its subjective certainty across contexts and species, both local circuit mechanisms and global network states need to be considered.

This symposium brings together four complementary approaches that collectively disentangle bottom-up sensory and arousal influences from top-down expectation and learning factors. First, Stijn Nuiten demonstrates that humans and mice alternate between discrete and persistent behavioral states and that pupil-linked arousal and GABAergic activity in sensory regions jointly facilitate task-optimal state probability. Next, Alessandro Toso reveals distributed, competing accumulators in premotor and motor cortices that encode both decision speed and confidence, highlighting a link between the balance of competing neural signals and subjective certainty. Sarah Tune then explores how pharmacologically altering the cortical excitation-to-inhibition balance affects the integration of auditory signals and prior expectations in perceptual decision-making and metacognition. Finally, Monja Froböse focuses on how pharmacological inhibition of cholinergic transmission shapes learning and decision-making under uncertainty, demonstrating that processes of probabilistic inference and adaptive choice behavior depend on specific neuromodulatory mechanisms when the brain is challenged by volatile environments.

By combining diverse pharmacological, behavioral and analytic approaches, this symposium offers convergent insights into the neural algorithms that govern how evidence is acquired, weighted, and experienced in perceptual decision-making, metacognition, and learning.

Behavioral responses to sensory inputs are highly variable, even upon repeated presentations of an identical stimulus. Traditional behavioral analyses (e.g., Signal Detection Theory) assume that this variability stems from uncorrelated noise whose average magnitude is static over time. However, a recent insight, afforded by generalized linear hidden Markov models (GLM-HMMs), is that humans and rodents alternate between discrete and persistent behavioral states during perceptual decision-making. For example, experimental trials can be clustered in states of engaged, disengaged, and biased decision-making strategies. In mice, the probability of being in an engaged behavioral state exhibits an inverted-U relationship to baseline pupil size (a proxy of tonic arousal), consistent with the Yerkes-Dodson Law. By analyzing behavioral, pupil, and neural data from mice (N=9; audio-visual change detection task) and humans (N=69; auditory detection task), the current study investigated 1) whether this relationship generalizes to human participants and 2) what neural mechanisms it is governed by. Mice and humans alternated between several discrete behavioral states and engaged behavioral state probability exhibited an inverted-U relationship with baseline pupil-linked arousal. In mice, preliminary neural analyses further suggest that for visual change detection, this relationship was mediated by pre-change V1 firing rates of putative GABAergic interneurons but not putative pyramidal neurons. These findings imply a general mechanism by which arousal dynamically modulates the cortical state of a primary sensory region to optimize perceptual decision-making. This study furthermore highlights an important insight for neuroscientific research: perception is governed by discrete and persistent states of altered sensory processing.

Neural circuit models of decision-making postulate that distinct populations of choice-selective neurons accumulate input signals supporting one versus another alternative. In this framework, the decision variable (DV) is encoded in a distributed (2-dimensional) fashion. It is unknown whether this distributed coding of DV predicts key aspects of decision-making. Here, we tested if distributed coding governs the confidence and reaction time (RT) associated with a choice.

Human participants (N=20) performed a visual task during magnetoencephalography (MEG) recordings. In each visual hemifield, a stream of 10 “samples” (circular gratings) of fluctuating contrasts was presented. Participants reported the “stronger” side (left vs. right) and their confidence in that choice (high vs. low) by button press. We estimated the impact of evidence fluctuations on choice and confidence and fit accumulator models to both features. Moreover, we trained decoders to predict choices from patterns of source-level MEG signals in the dorsal premotor and primary motor cortex (PMd/M1) of each hemisphere.

Choice and confidence reports were biased by evidence fluctuations across all samples. For choice this impact was symmetric for “chosen” and “unchosen” streams but dominated by the chosen stream for confidence. Behavior was explained by a model of two competing leaky accumulators, racing to hit a collapsing bound. The output of PMd/M1 choice decoders built up during decision formation. Critically, RTs and confidence reports depended on decoder outputs from both hemispheres.

Our results establish a distributed representation of decision states in the brain, whereby the balance between competing neural accumulators shapes RT and confidence.

Perceptual decisions are rarely dictated by sensory input alone: prior expectations and internal neural states also influence how sensory evidence is encoded, weighted, and integrated over time. In my talk, I will present on two EEG studies which leveraged an adapted click train paradigm [Brunton et al. (2013)] to investigate how these processes jointly shape adaptive yet sometimes suboptimal auditory decision-making and metacognition.

Participants listened to 1-sec trains of 20 clicks, randomly presented to either left or right ear, and judged which side received more clicks. Prior expectations were manipulated via probabilistic visual cues: For half of the trials, an informative cue indicated which ear was the more likely target side in a given trial. I will show how these expectations are integrated with ongoing sensory evidence, and how variability in pre- and peri-stimulus neural dynamics relates to decision performance and confidence.

In the first study (N=32, 18–33 yrs), prior expectations systematically biased perceptual decisions. Here, we asked in how far these biases can be explained by cue-driven as well as spontaneous fluctuations in neural dynamics modulating the analysis and integration of sensory evidence.

In our second study (N=14, 18–41 yrs), we pharmacologically perturbed cortex-wide neural dynamics to more directly probe how fluctutations in excitation-inhibition balance interact with prior expectations to guide both perceptual judgments and confidence.

Collectively, I will show how global neural states and prior expectations shape the computations underlying perceptual decisions and metacognition.

The neuromodulator acetylcholine has been suggested to both govern the rate of learning under uncertainty and to modulate neural circuits relevant for reward-guided choices irrespective of learning. To investigate the role of acetylcholine in decision making under uncertainty, we administered the muscarinic M1 receptor antagonist biperiden to participants performing a reward-guided decision-making task in the presence and absence of learning, and in volatile versus stable learning environments. Specifically, participants performed two tasks that involved choices between options characterized by two attributes, reward probability and magnitude. In the gambling task, both attributes were explicitly provided, whereas in the learning task, reward probabilities had to be inferred from past experience. In addition, uncertainty was manipulated within the learning task by inclusion of a stable phase with fixed reward contingencies, and a volatile phase with frequent contingency reversals. Healthy male participants (n = 43) performed these tasks after administration of biperiden (4 mg) in a within-subjects, placebo-controlled design. Biperiden did not affect decision making in the gambling task, where no learning was required. However, in the learning task, biperiden reduced the sensitivity to the learnt reward probabilities. Notably, this was primarily driven by choices under higher uncertainty in the volatile phase. Using reinforcement learning models, we show that the change in behaviour was caused by noisier estimates of probabilities resulting from maladaptively increased learning rates under biperiden. Together, these findings suggest that muscarinic acetylcholine transmission is involved in controlling learning in highly uncertain contexts, when the demand for carefully calibrated adjustments is highest.