Abstract
Psychophysical methods are a cornerstone of psychology and neuroscience, used to quantify behavior through controlled experiments and signal detection theory. However, traditional methods are often tedious and time-consuming. Continuous psychophysics offers a faster alternative by using fluid behavioral adjustments to dynamic stimuli, but previous analysis methods yielded perceptual thresholds an order of magnitude larger than traditional tests. We introduce a Bayesian inverse optimal control framework that accounts for action variability, internal costs, and subjective beliefs. This approach recovers accurate perceptual thresholds and provides insights into the intrinsic costs of behavior, demonstrating that perception is best understood as part of an active control loop.
Introduction: The Accuracy Gap in Continuous Psychophysics
Traditional psychophysics is the "gold standard" for measuring sensory thresholds, but it relies on repetitive, discrete trials (e.g., "Yes/No" or "Left/Right" choices). Continuous psychophysics replaces this rigid structure with naturalistic tracking tasks, where participants follow a moving target with their eyes, hands, or a mouse. While significantly faster and more engaging, early attempts to analyze this data resulted in perceptual uncertainty estimates ten times higher than those from traditional methods.
This "accuracy gap" suggested that previous models—often simple Kalman Filters—were missing critical components of how humans actually perform tracking tasks. By misattributing motor noise and effort-saving strategies to perceptual failure, these models created a "junk drawer" effect that inflated sensory thresholds.
Methods: A Bayesian Inverse Optimal Control Framework
The researchers propose that tracking is not merely a perceptual task but a Partially Observable Markov Decision Process (POMDP). Within a Linear-Quadratic-Gaussian (LQG) framework, they applied Bayesian Inverse Optimal Control to infer internal parameters from raw behavioral data. This allows the model to separate the noise inherent in the sensory system from the noise in the motor system and the subjective costs of movement.
The framework evaluates four nested models of increasing complexity:
- Ideal Observer: Models only perceptual noise, assuming the actor is otherwise perfect.
- Optimal Actor: Incorporates motor (action) variability, acknowledging that our movements are inherently noisy.
- Bounded Actor: Includes internal "behavioral costs" representing the effort of movement; humans often choose to be slightly less accurate to save energy.
- Subjective Actor: Accounts for the participant's subjective beliefs about stimulus dynamics, such as assuming a target has momentum even when it follows a random walk.
Key Insights: Beyond the "Junk Drawer"
A major finding is that much of the error in continuous tracking is actually caused by motor noise and a desire to minimize effort. When models ignore these factors, the perceptual noise estimate becomes a catch-all for every other source of variability. By explicitly modeling the cost of action, the researchers found that humans track imperfectly not because they cannot see the target, but because high-velocity corrections are cognitively and physically "expensive."
Furthermore, participants often harbor "subjective physics"—internal beliefs about the target's movement that differ from the actual experimental parameters. For instance, participants might assume a target has velocity-based momentum. Modeling these beliefs significantly improved the fit to human data across mouse, hand, and gaze tracking tasks, even in non-human primates like macaques and marmosets.
Results and Efficiency
The new framework successfully closed the accuracy gap. While previous methods overestimated thresholds by 1000%, the LQG-based model produced estimates within ~1.36x of traditional results. This brings continuous psychophysics into alignment with the gold-standard 2-interval forced-choice (2IFC) results.
Beyond accuracy, the framework offers a massive leap in efficiency. The study suggests that just 2 minutes of continuous tracking can yield perceptual estimates as stable as 20 minutes of traditional trial-based testing. This 10-fold increase in data collection efficiency could revolutionize clinical and experimental settings where time is a limiting factor.
Conclusion
By treating perception as an active control loop that includes motor constraints and subjective costs, this framework makes continuous psychophysics a viable, high-accuracy tool for sensory research. It underscores the importance of considering the "actor" within the "observer" when studying human perception, proving that our sensory thresholds are deeply intertwined with how we choose to move.
Summary updated