Perceptual Decision Making of Nonequilibrium Fluctuations

How does the brain recover a weak signal that is submerged in intense stochastic fluctuations to make fast yet accurate choices? We recast perceptual decision-making as the inference of a nonzero drift (v) in the presence of large diffusivity (D): the observer must determine the direction of motion when trajectories are dominated by diffusion. A concrete analogy is reading wind while hunting: turbulent gusts scramble moment-to-moment cues, yet a subtle, consistent drift in air motion carries actionable information about wind direction. In this perspective, the core problem is signal discovery under diffusion—extracting the sign and magnitude of a weak drift from time-resolved fluctuations. This framing contrasts with the traditional random-dot motion (RDM) paradigm, where “coherence” (the percentage of dots moving consistently) is an effective but unitless control of difficulty. Although 50% coherence is intuitively “more signal” than 10%, its relationship to the quantity of evidence is ambiguous: is it five times more, or something else? Because coherence does not specify the physical statistics generating samples, an ideal-observer analysis (e.g., a Sequential Probability Ratio Test, SPRT) cannot be uniquely grounded in first principles. By instead specifying a physics-based stimulus with known parameters, this thesis makes the evidence quantified and the ideal-observer well-posed. We generated visual motion stimuli from a drifted Brownian process. To index the distance from equilibrium, we employed an interpretable nonequilibrium measure (entropy production Σ) proportional to drift–diffusion contrast, which increases as directional drive overwhelms diffusivity. In this generative setting, the momentary log-likelihood ratio (LLR) for direction decisions and the optimal stopping boundaries of the SPRT are derived analytically from (v, D), providing a physics-grounded benchmark for ideal performance. Across three behavioral experiments, we asked: i) whether human observers (N = 67) detect and exploit graded nonequilibrium dynamics; ii) how closely their choices approach an ideal-observer benchmark; iii) how evidence integration adapts as Σ varies; and iv) whether such adaptation depends on task structure and the spatiotemporal layout of the stimuli. Results showed that stimulus dynamics (Σ, v, D) robustly shaped decision metrics, demonstrating that observers are indeed sensitive to graded changes in Σ. An analytical SPRT captured these effects and quantified deviations from ideal performance. Complementarily, an Evidence Integration Model (EIM) fitted to the data revealed a systematic adjustment of the temporal integration window with Σ: in each trial, observers assigned greater weight to the most recent portion of the trajectory (a recency effect) whose strength scaled with Σ. Observers were also sensitive to salient changes in trajectory directionality, consistent with adaptive weighting under nonstationary drift. Crucially, these effects were stronger in a blocked design—where Σ was held constant within blocks—than in an intermixed design, where Σ varied from trial to trial, indicating that stable nonequilibrium statistics facilitate calibration of integration timescales. Finally, sensitivity to the nonequilibrium structure was modulated not only by the physical parameters (v, D) but also by the spatial and temporal layout of the stimuli. Overall, by embedding perceptual evidence in a physics-based process that specifies its quantity, this work refines the characterization of variables that govern perceptual decisions and clarifies the temporal dynamics underlying efficient sensory evidence integration. It shows that when evidence is measured—rather than merely manipulated—ideal-observer analyses become principled, enabling precise tests of how the brain detects weak directional signals under high diffusion.