Late Integration of Prior Expectations During Precision Weighted Perceptual Decisions

Through two preregistered EEG experiments, this study demonstrates that precision-weighted integration of prior expectations and sensory evidence in visual motion decisions occurs primarily during late response planning stages rather than early sensory processing, thereby challenging predictive processing accounts.

The Gist

The Big Question: How Does Your Brain Guess What's Coming?

Imagine you are walking through a foggy forest. You hear a rustle in the bushes. Is it a bear? A deer? Or just the wind?

To make a decision, your brain does two things:

1. Looks at the evidence: How clear is the sound? (Is the fog thick or thin?)
2. Checks its memory: What usually happens in this forest? (Are bears common here, or is it usually just wind?)

Scientists call this Bayesian Inference. It's the idea that your brain mixes "what you see" with "what you expect" to make the best guess possible. The smarter your brain is, the more it weighs the evidence based on how reliable it is. If the fog is thick (bad evidence), you rely more on your memory. If the fog is thin (good evidence), you trust your eyes more.

The Mystery: We know humans are really good at doing this math in their heads. But where in the brain does this magic happen? Does the brain change how it sees the world based on expectations (Early Processing), or does it change how it decides what to do after seeing the world (Late Processing)?

The Experiment: A Game of Moving Dots

The researchers set up a game to find out.

• The Task: Participants sat in front of a screen watching a cloud of white dots moving around. Sometimes the dots moved together in a clear direction (like a school of fish). Sometimes they moved chaotically (like a swarm of confused bees).
• The Twist: Before the dots appeared, the participants learned a "rule" for the current round.
  • Rule A: "The dots usually go Up." (Strong Expectation)
  • Rule B: "The dots go everywhere randomly." (No Expectation)
• The Goal: The participants had to guess the direction the dots were moving.

They measured two things:

1. How well they did: Did they guess correctly?
2. What their brains were doing: They used EEG (a cap with sensors) to listen to the brain's electrical signals in real-time.

The Findings: The "Late Bloomer"

Here is what they discovered, broken down simply:

1. The Behavior: We Are Smart Statisticians

The participants were excellent at the game. When the dots were hard to see (chaotic), they leaned heavily on the "rule" they learned. When the dots were easy to see (clear movement), they ignored the rule and trusted their eyes.

• Analogy: It's like driving in the rain. If the windshield wipers are working well (clear dots), you drive by looking at the road. If the wipers are broken and it's pouring (blurry dots), you slow down and rely on your memory of the route.

2. The Brain Signal: The "Decision Meter" (CPP)

The researchers looked at a specific brain signal called the CPP (Central Parietal Positivity). Think of this as a decision meter that fills up like a cup of water.

• When the dots were clear, the cup filled up fast.
• When the dots were blurry, the cup filled up slowly.
• The Surprise: When the participants had a strong "rule" (expectation), the cup didn't fill up faster. Instead, it started with a head start (it was already partially full because they expected that direction).
• Meaning: The expectation didn't make their eyes see better; it just gave them a head start on the decision.

3. The Big Reveal: Expectations Arrive Late

This is the most important part. The researchers used a special technique to "decode" what the brain was thinking about the direction of the dots.

• Early Stage (Seeing the dots): The brain's representation of the dots was purely based on what was actually on the screen. If the dots were blurry, the brain's "picture" was blurry. Expectations did not change the picture.
• Late Stage (Planning the answer): Just before the participant moved their hand to answer, the brain's picture changed! If they had a strong expectation, the brain's representation of the dots became sharper and more accurate.
• Analogy: Imagine you are trying to remember a song.
  • Early stage: You hear a few notes. Your brain records exactly what it hears.
  • Late stage: As you are about to sing the song, your brain "fills in the gaps" based on what you think the song should be. The expectation helps you sing it better, but it didn't change what you heard in the first place.

Why This Matters

For a long time, many scientists believed in Predictive Processing. This theory suggests that your brain is like a projector that paints the world onto your eyes before you even see it. They thought expectations changed how we see things instantly.

This paper says: "Not so fast."

The study shows that while our brains are amazing at using expectations to make decisions, they don't use them to change the raw sensory input. Instead, they use expectations to polish the decision right before we act.

• Old View: Expectations act like sunglasses that tint the world before you look at it.
• New View: Expectations act like a coach standing next to you, whispering advice after you've seen the play, helping you make the right move at the last second.

The Takeaway

Your brain is a brilliant statistician that knows how to mix "what is happening" with "what usually happens." But it keeps these two things separate for a while. It first records the raw data, and then, just as you are about to act, it uses your past experience to sharpen your focus and guide your hand.

In short: We don't see what we expect; we decide based on what we expect.

Technical Summary

1. Problem Statement

Adaptive perceptual decision-making requires the brain to combine uncertain sensory evidence (likelihoods) with prior expectations in a precision-weighted manner, a process described by Bayesian inference. While behavioral studies confirm that humans perform this integration optimally, the neural mechanisms and temporal dynamics of how the brain represents and combines these priors and likelihoods remain unclear.

Two competing theoretical frameworks exist:

• Predictive Processing: Suggests that priors modulate early sensory processing, sharpening or biasing neural representations of expected stimuli immediately upon (or even before) stimulus onset.
• Accumulate-to-Bound Models: Suggest that priors act as a response bias (shifting the starting point of evidence accumulation) or alter the rate of accumulation later in the decision process, without necessarily altering early sensory encoding.

This study investigates whether precision-weighted integration occurs early (sensory encoding) or late (decision formation/action planning) and whether the precision of the prior scales with the precision of the sensory evidence at the neural level.

2. Methodology

Experimental Design
The authors conducted two preregistered experiments using a random-dot motion estimation task:

• Experiment 1 (Behavioral): $N=40$ participants.
• Experiment 2 (EEG + Behavioral): $N=40$ participants.
• Task: Participants estimated the direction of a coherent motion signal embedded in noise.
• Manipulations:
  • Sensory Precision (Likelihood): Varied via motion coherence ($\kappa_s$). In Exp 2, three levels were used: Low (0.2), Medium (0.4), and High (2.0).
  • Prior Precision: Manipulated block-wise by varying the probability distribution of target motion directions. Priors were Uninformative (Uniform, $\kappa_p=0$), Weakly Informative ($\kappa_p=5$), or Strongly Informative ($\kappa_p=10$).
• Procedure: Participants learned the prior statistics for specific blocks and reported motion direction using a 360-degree on-screen dial.

Neural Analysis (Experiment 2)

• EEG Recording: 64-channel EEG recorded at 1024 Hz.
• Univariate Analysis: Examined the Centro-Parietal Positivity (CPP), a well-established neural correlate of evidence accumulation. Amplitude and slope were analyzed relative to stimulus precision and prior precision.
• Multivariate Analysis (Inverted-Encoding Models - IEM):
  • Trained linear models to map EEG sensor activity to 16 hypothetical motion-tuned channels.
  • Inverted the model to recover population tuning profiles from held-out data.
  • Metrics: Decoding accuracy (fidelity of motion representation) and Neural Bias (shift in represented direction toward the prior).
  • Temporal Generalization: Trained models on the stimulus observation period (250–750 ms) and tested them on the response period to distinguish between stimulus-driven and response-driven activity.

3. Key Contributions

1. Temporal Dissociation of Priors and Likelihoods: The study provides high-temporal-resolution evidence that while sensory precision affects neural representations immediately (during stimulus encoding), prior precision only influences neural representations during late stages (response planning/execution).
2. Challenge to Predictive Processing: The findings contradict influential predictive processing theories that posit priors modulate early sensory representations. Instead, the data supports a model where priors influence the decision process (e.g., starting point of accumulation) rather than the sensory encoding itself.
3. Absence of Neural Bias: Despite robust behavioral biases toward the prior mean, the neural representation of the motion direction remained veridical (faithful to the actual stimulus) throughout the trial. This suggests behavioral biases arise during response preparation, not sensory encoding.
4. Precision-Weighted Neural Fidelity: The study demonstrates that the brain maintains a "memory" of the stimulus with higher fidelity when the prior is informative, but this enhancement is specific to the decision/action phase.

4. Key Results

Behavioral Findings (Both Experiments)

• Precision-Weighted Inference: Response accuracy improved with higher sensory coherence and stronger priors.
• Interaction: The benefit of a strong prior was most pronounced when sensory evidence was weak (low coherence), confirming Bayesian-like integration.
• Bias: Participants showed strong attractive biases toward the prior mean, which increased with prior strength and decreased with sensory precision.

Neural Findings (Experiment 2)

• CPP (Evidence Accumulation):
  • Stimulus Precision: Higher coherence led to steeper CPP slopes (faster accumulation).
  • Prior Precision: Stronger priors resulted in lower overall CPP amplitude and earlier peaks, consistent with a shift in the starting point of accumulation (less evidence needed to reach the bound), but did not alter the rate of accumulation (slope).
• Multivariate Decoding (Motion Representations):
  • Stimulus Period: Decoding accuracy was sensitive to sensory precision (high coherence = high accuracy) but insensitive to prior precision.
  • Response Period: Decoding accuracy was sensitive to both sensory and prior precision. Stronger priors led to more accurate decoding of motion direction during response planning.
  • Temporal Generalization: Models trained on stimulus observation data successfully predicted the effect of prior precision during the response period. This indicates that the enhanced representation under strong priors is not merely a motor artifact but reflects a sustained, stimulus-correlated neural state.
• Neural Bias: No systematic shift in the decoded motion direction was observed toward the prior mean at any timepoint. The neural representation tracked the veridical stimulus, not the Bayesian posterior.

5. Significance

This work advances the understanding of the neural architecture of perceptual decision-making by establishing a clear temporal hierarchy:

1. Early Stage (Sensory Encoding): Neural representations are driven strictly by the reliability of the incoming sensory evidence. Priors do not "pre-tune" these early sensory signals in this continuous estimation context.
2. Late Stage (Decision/Action): Priors are integrated to enhance the fidelity of the stimulus representation and guide the decision process. This integration manifests as a shift in the starting point of evidence accumulation and an enhancement of the stimulus representation during response preparation.

These findings challenge the notion that the brain constantly "hallucinates" expected stimuli to reduce sensory load (as suggested by some predictive processing accounts). Instead, they suggest a more conservative strategy where the brain waits for sensory evidence, accumulates it, and only then applies prior knowledge to optimize the final decision and action. This has implications for how we model neural computation in AI and understand disorders involving sensory integration.