The textbook model of Bayesian perception is one of attraction. If you have a prior belief about the world—for instance, that most lines in our environment are vertical or horizontal—and you are presented with a noisy, ambiguous stimulus, your brain should theoretically pull your perception toward that prior expectation.
Yet, decades of psychophysical experiments reveal a glaring paradox: human perception frequently exhibits repulsive biases. When individuals adapt to a specific stimulus, their subsequent estimates shift away from it. When they make a categorical choice, their fine-grained perceptions are repelled from the decision boundary. Even when estimating basic sensory features, their judgments are often pushed away from the most common environmental states.
How can a rational, Bayesian brain produce systematic deviations away from its own expectations and high-precision anchors? A growing body of research resolves this paradox by looking under the hood of how the brain encodes and decodes information, revealing that repulsive biases are not cognitive bugs, but the mathematical consequences of efficient resource allocation and conditional inference.
Efficient Coding and Asymmetric Likelihoods
To understand why perception can be repelled from a prior, we must look at how neurons represent information. The principle of efficient coding dictates that the brain should not waste precious metabolic resources encoding all stimuli equally. Instead, it prioritizes precision where it matters most, allocating more neural resources (and thus higher encoding precision) to the most frequently encountered stimuli.
This creates an immediate tension. As noted by Prat-Carrabin et al., the final bias in a perceptual system emerges from a tug-of-war between two opposing forces: a prior-driven force pulling estimates toward high-probability regions, and an encoding-driven force pushing estimates away from regions of high precision.
Standard Bayesian (Uniform Precision):
Prior Peak (Attraction) <=========== Percept
Efficient Coding (Variable Precision):
Prior Peak (Attraction) <=========== Percept ===========> Asymmetric Likelihood (Repulsion)
Why does high precision cause repulsion? In a classic paper, Wei et al. demonstrated that because encoding precision varies across a stimulus space, the internal likelihood function generated by a stimulus is fundamentally asymmetric. When a stimulus appears slightly off-peak from a high-precision zone, the side of the likelihood function closer to the high-precision zone drops off much faster than the side facing away.
When the brain decodes this skewed likelihood, the asymmetry pushes the peak of the posterior distribution away from the high-precision region. Internal neural noise inherently drives this repulsive bias. Morais et al. generalized this phenomenon, proving that anti-Bayesian repulsive biases are a mathematical inevitability across a broad family of power-law efficient neural codes whenever less common stimuli are encoded with relatively higher precision than a standard flat-precision observer would assume.
To unify these seemingly contradictory attractive and repulsive behaviors, Xue-Xin Wei formulated the Prior/Precision ratio rule (Wei, 2023). This universal rule dictates that the direction of perceptual bias depends entirely on the relative spatial rates of change between prior beliefs and encoding precision. If external stimulus noise is high, the attractive pull of the prior dominates; if internal neural noise dominates, the repulsive push of the asymmetric likelihood wins. This exact mechanism explains why human probability distortion—the tendency to overweight rare events and underweight likely ones—is actually a rational feature of decoding noisy representations where neural resources are prioritized at extreme boundaries, as shown by Tong et al..
Adaptation as Dynamic Resource Reallocation
This efficient coding framework provides an elegant explanation for sensory adaptation. Traditional models struggled to explain why staring at a stimulus causes subsequent percepts to be repelled away from it.
Stocker et al. reconciled this within a Bayesian framework by demonstrating that adaptation does not alter our long-term prior expectations. Instead, adaptation is a dynamic, short-term reallocation of measurement resources that spikes encoding precision directly around the adaptor. By keeping the prior fixed and modulating the local likelihood precision, the measurement stage naturally generates a powerful repulsive force that pushes subsequent estimates away from the recent stimulus.
This theory yields a clear, testable prediction: if repulsive adaptation biases are driven by the brain trying to navigate internal uncertainty, then individuals with noisier sensory representations should exhibit distinct bias patterns. Carter et al. confirmed this empirically, demonstrating that individuals with lower baseline sensory precision exhibit significantly stronger repulsive adaptation biases. When internal representations are highly uncertain, the brain relies more heavily on these adaptive shifts in likelihood precision to interpret the world.
Selection Biases from Causal Inference and Decisions
Repulsion does not only happen at the hardware level of neural encoding; it also emerges at the software level of decision-making and categorization.
Consider multisensory perception, such as combining what we see and what we hear. The brain does not blindly average these signals. As established by Körding et al., the brain performs causal inference, calculating the probability that the two signals arose from a single shared source or two independent causes.
When the brain infers that a visual and an auditory cue stem from independent causes, the resulting spatial estimates often exhibit a negative, repulsive bias—appearing further apart than they actually are. The authors of Causal Inference in Multisensory Perception explained that this is a statistical selection bias. Because of internal sensory noise, the distance between two cues fluctuates from trial to trial. The brain is highly likely to categorize the signals as "independent" on precisely those trials where internal noise happened to artificially inflate the perceived distance between them. Slicing the data to look only at "independent cause" trials isolates a subset of data that is statistically guaranteed to show a repulsive bias.
A closely related form of repulsion occurs when a fine-grained estimation is preceded by a categorical choice. If you must first decide whether a line is tilted left or right, and then estimate its exact angle, your final estimate will be systematically repelled away from the decision boundary.
To explain this, Stocker et al. proposed a conditional Bayesian observer model. They argued that the human brain sacrifices absolute mathematical optimality in favor of internal self-consistency. Once a categorical decision is made, the brain reduces cognitive dissonance by discarding all hypotheses that contradict that choice.
Full Posterior Distribution:
|
XXXXX|XXXXX <- Decision Boundary
|
Conditional Posterior (After choosing "Right"):
|
|XXXXX <- Left side discarded; mean shifts right (Repulsion)
|
By conditioning the subsequent estimation strictly on the chosen category, the brain effectively truncates the probability distribution at the decision boundary. Slicing away the unchosen side of the distribution forces the mean of the remaining distribution to shift sharply away from the boundary, creating a powerful, post-decision repulsive bias.
Biological Constraints on Ideal Inference
While these models demonstrate the hidden rationality behind repulsive biases, human perception is ultimately bound by biological constraints. Ideal observer models assume a perfect statistical decoder, but real brains rely on population codes that can be suboptimal.
For instance, Putzeys et al. demonstrated that human population decoding suffers from a "blurry" read-out process that fails to completely filter out task-irrelevant background noise, introducing unexpected systematic biases. Similarly, Fetsch et al. noted that even when animals dynamically reweight cues based on reliability, they still maintain fixed, idiosyncratic biases toward specific sensory modalities that defy pure statistical optimality.
Ultimately, repulsive biases reveal that perception cannot be understood by looking at priors or sensory inputs in isolation. Whether driven by the asymmetric push of efficient neural allocation, the statistical filtering of causal inference, or the self-consistency constraints of a conditional choice, repulsion is the signature of a brain dynamically optimizing its limited resources to make sense of an uncertain world.
Key Foundational Work
Sensory Adaptation within a Bayesian Framework for Perception (2005) — contributed 2 concept elements to this direction.
Causal inference in sensorimotor integration (2006) — contributed 1 concept elements to this direction.