Structural and contextual biases interact to shape duration perception

This study demonstrates that human duration perception is shaped by the interaction of structural biases, such as the auditory dominance over visual stimuli, and contextual biases driven by environmental statistics, revealing that the brain employs a combination of Bayesian inference and rescaling mechanisms to normalize temporal representations.

The Gist

The Big Idea: How Your Brain Tells Time

Imagine your brain is a master chef trying to guess how long a soup has been simmering. The chef doesn't have a stopwatch; they have to rely on two things:

1. The Recipe (Structural Priors): Rules that are always true, like "sound travels faster than sight" or "sounds usually feel longer than lights."
2. The Kitchen Context (Contextual Priors): The specific situation right now. Is the kitchen full of short, quick stirs? Or long, slow simmers?

This paper asks: How does the chef mix these two ingredients to decide if the soup is ready?

The researchers found that the brain does something surprisingly clever. It doesn't just follow the recipe or just look at the context. It does both, but it also performs a mental "stretch and shrink" trick to make sure the two ingredients can be compared fairly.

---

The Experiment: The "Which Lasted Longer?" Game

The researchers set up a game for 30 people.

• The Game: Two sounds or two flashes of light (or one of each) would play one after the other.
• The Task: The player had to press a button to say, "Which one lasted longer?"
• The Twist: The researchers changed the "menu" (the context). Sometimes, the game only used short durations (like quick beeps). Other times, it used long durations (like slow chimes).

What they expected:
Usually, if you are used to hearing short beeps, a medium beep feels long. If you are used to long chimes, that same medium beep feels short. This is called the Central Tendency (or "Regression to the Mean"). It's like your brain saying, "This is average for this group, so I'll guess it's average."

What they found:
The brain did the opposite of what simple guessing would predict. When the brain saw a medium beep in a group of long chimes, it didn't just think "it's average." It actually thought, "Wait, this is shorter than the others!"

The Solution: The "Rubber Band" Analogy

To explain this weird result, the researchers proposed a model with two main steps:

1. The "Two Separate Notebooks" (Distinct Representations)

The brain doesn't mix all the short and long sounds into one big pile. Instead, it keeps two separate notebooks.

• Notebook A: Contains the rules for the "Short" group.
• Notebook B: Contains the rules for the "Long" group.
  When a sound plays, the brain knows exactly which notebook to open. It doesn't get confused by the other group.

2. The "Rubber Band Stretch" (Rescaling)

This is the most important discovery. Before the brain compares the two sounds, it stretches or shrinks the mental ruler.

Imagine this:
You are comparing the height of two people.

• Person A is standing in a room where everyone is 6 feet tall.
• Person B is standing in a room where everyone is 5 feet tall.

If you just look at them, Person A looks tall and Person B looks short. But to compare them fairly, your brain grabs a rubber band (the rescaling mechanism). It stretches the "5-foot room" and shrinks the "6-foot room" so that both people are standing on the same flat floor.

In the experiment, the brain realized: "Oh, this sound came from the 'Long' group, so I need to shrink my perception of it to compare it fairly to the 'Short' group." This rescaling fights against the usual "guessing the average" instinct.

The "Sound vs. Sight" Bias (Structural Priors)

The study also looked at a built-in bias in our brains: We always think sounds last longer than lights of the same length.

• If a light flashes for 1 second and a beep lasts for 1 second, you will almost always say the beep felt longer.
• The researchers found this is a hard-wired rule (a "structural prior") that everyone has, though some people have it more strongly than others. It's like having a permanent filter on your glasses that makes sound "stretch" more than sight.

Why Does This Matter?

This paper changes how we think about time perception.

• Old Idea: Our brain is a passive recorder that gets confused by context.
• New Idea: Our brain is an active editor. It constantly re-calibrates its internal clock to fit the current environment. It creates a "common language" for time so it can make fair comparisons, even when the environment changes.

Summary in a Nutshell

Your brain is like a smart translator.

1. It knows that sounds naturally feel longer than lights (Structural Bias).
2. It keeps separate dictionaries for different groups of time (Short vs. Long).
3. Before making a decision, it stretches or shrinks the mental ruler (Rescaling) so that a "long" sound and a "short" sound can be compared on the same scale.

This explains why our perception of time is so flexible and why we can adapt so quickly to different environments, from a fast-paced video game to a slow, quiet library.

Technical Summary

1. Problem Statement

Human time perception is highly variable, influenced by both structural priors (stable, idiosyncratic brain constraints effective in all situations, such as modality-specific processing differences) and contextual priors (adaptive constraints based on the statistical distribution of stimuli in a specific environment).

• The Gap: While contextual priors (e.g., regression to the mean based on stimulus distributions) are well-studied, structural priors remain underexplored. Furthermore, no unified computational model currently integrates how these two distinct sources of bias interact during temporal discrimination.
• The Question: How does the brain combine stable structural biases (e.g., auditory stimuli perceived as longer than visual ones) with dynamic contextual biases (e.g., adaptation to short vs. long duration distributions) to form a duration estimate?

2. Methodology

Experimental Design

• Task: A two-interval duration discrimination task. Participants judged which of two sequentially presented stimuli (S1 and S2) lasted longer.
• Stimuli: Visual (pink noise annuli) and Auditory (pink noise bursts).
• Conditions:
  • Modality: Both visual (VV), both auditory (AA), or mixed (AV, VA).
  • Contextual Priors: Manipulated via two distinct duration distributions:
    • Short Distribution: Median 0.35s (range 0.2–0.6s).
    • Long Distribution: Median 0.6s (range 0.35–1.033s).
  • Combinations: S1 and S2 were drawn from either the same or different distributions (e.g., Short-Long vs. Long-Short), creating 8 main conditions plus 4 baseline conditions (where S1 was fixed to cancel out contextual priors).
• Participants: 30 healthy adults.

Computational Modeling

The authors developed and compared five Bayesian models to fit individual behavioral data. The models varied in three key components:

1. Distribution Representation:
   • 1distrib: Assumes a single unified prior distribution encompassing all possible durations.
   • 2distrib: Assumes the brain maintains separate prior representations for the distribution of S1 and the distribution of S2.
2. Rescaling Mechanism:
   • No Rescaling: Standard Bayesian inference where estimates regress toward the mean of the specific distribution (central tendency).
   • Rescaling: A mechanism where the perceptual system realigns the two distributions onto a common internal axis (relative coding) before comparison. This predicts an effect opposite to the central tendency (e.g., a stimulus drawn from a "long" distribution might be perceived as shorter relative to a stimulus from a "short" distribution if the system normalizes the ranges).
3. Structural Prior: A parameter ($\delta$) accounting for the systematic overestimation of auditory vs. visual durations.
4. Order Effect: A parameter ($\alpha$) scaling the uncertainty of the first stimulus (S1) to account for memory decay.

Model Fitting: Parameters were optimized using Bayesian Adaptive Direct Search (PyBADS) to minimize the Negative Log Likelihood (NLL). Models were compared using the Akaike Information Criterion (AIC).

3. Key Contributions

• Unified Framework: The study provides the first computational model that simultaneously quantifies structural and contextual priors in a single discrimination task.
• Discovery of Dual Mechanisms: It demonstrates that duration discrimination is not driven solely by Bayesian regression-to-the-mean. Instead, it requires a combination of Bayesian inference (contextual bias) and rescaling (normalization).
• Separate Distribution Encoding: Evidence suggests the brain maintains distinct representations for the statistical distributions of the first and second intervals, rather than merging them into a single global prior.
• Quantification of Structural Bias: The study isolates and quantifies the structural prior (auditory overestimation) independently of contextual effects.

4. Key Results

Model Comparison

• The "rescaling+2distrib" model provided the best fit to the data (lowest AIC) for the majority of participants.
• This model posits that the brain:
  1. Maintains separate priors for S1 and S2 distributions.
  2. Applies a rescaling step to realign these distributions onto a common internal scale before comparison.
• Pure Bayesian models ("2distrib" without rescaling) failed to capture the data, particularly for overlapping durations where the rescaling effect opposes the central tendency effect.

Parameter Estimates

• Rescaling ($c$): The rescaling parameter was estimated at $c \approx 0.65$, indicating a strong but partial realignment of the perceptual spaces. This suggests the brain normalizes temporal representations to adapt to environmental statistics.
• Structural Prior ($\delta$): Participants consistently overestimated auditory durations compared to visual ones by an average of 7% (approx. 31 ms for a 450 ms stimulus). This was confirmed by a significant correlation between the parametric model estimate and a non-parametric behavioral measure.
• Uncertainty: Auditory stimuli had significantly lower uncertainty ($\sigma_A \approx 0.29$) compared to visual stimuli ($\sigma_V \approx 0.37$).
• Order Effect: The uncertainty of the first stimulus was 47% larger than the second ($\alpha \approx 1.47$), supporting the hypothesis that S1 suffers from memory decay or increased noise during the delay.

Parameter Recovery

Simulations confirmed the robustness of the best-fitting model, with high correlations between true and recovered parameters, though slight trade-offs were observed between the rescaling parameter ($c$) and the structural prior ($\delta$).

5. Significance and Implications

• Challenging Pure Bayesian Accounts: The findings challenge the view that temporal perception is solely governed by Bayesian regression to the mean. The presence of a rescaling mechanism suggests the brain employs a "relative coding" strategy (comparing stimuli relative to their local context) which can override or counteract central tendency effects.
• Adaptive Normalization: The rescaling mechanism implies that the brain actively normalizes temporal representations to adapt to the range of durations encountered, similar to luminance adaptation in vision. This allows for optimal discrimination within a specific context.
• Neural Mechanisms: The dissociation between structural (stable) and contextual (adaptive) priors suggests they rely on distinct neural mechanisms. Structural priors may reflect fixed connectivity or intrinsic properties of sensory pathways, while contextual priors and rescaling likely involve dynamic, higher-order predictive coding mechanisms.
• Clinical Relevance: By quantifying individual differences in these biases, the framework offers a potential tool for investigating temporal processing deficits in psychiatric or neurological conditions where predictive mechanisms are disrupted.

In summary, the paper establishes that human time perception is a complex interplay where the brain simultaneously maintains stable structural biases (modality differences) and dynamically rescales contextual information to optimize discrimination, moving beyond simple Bayesian integration.