On the inseparability of the prior and neural resources in behavioural bias

This paper demonstrates that behavioral biases in sensory decision-making arise not from separable prior expectations and neural resources, but from the direct embedding of prior information into neural tuning curves, which is supported by re-analyzed V1 physiology and allows for efficient, Bayes-optimal inference.

The Gist

The Big Question: How Do We See the World?

Imagine your brain is a detective trying to solve a mystery based on blurry clues (sensory input). To solve the mystery, the detective uses two things:

1. The Clues: What the eyes actually see right now (e.g., a slightly tilted line).
2. The Experience: What the detective expects to see based on past experience (e.g., "Lines in this room are usually vertical").

In the past, scientists thought the brain handled these two things separately. They thought the brain had a "clue-gathering machine" (the eyes) and a separate "expectation calculator" (the brain's memory) that combined them later.

This paper argues that the brain doesn't do it that way. Instead, the "clue-gathering machine" is built with the expectations already baked into its gears.

---

The Analogy: The Weather-Adjusted Umbrella Factory

Imagine a factory that makes umbrellas.

• The Old Way (Separate): The factory makes all umbrellas exactly the same size. Then, a separate manager looks at the weather forecast. If it's going to rain hard, the manager tells the factory, "Make the umbrellas bigger!" The factory and the manager are two different steps.
• The New Way (Integrated): The factory is built so that the machines automatically make bigger umbrellas in the rainy season and smaller ones in the dry season. The factory's design is the weather forecast. You don't need a separate manager to tell them what to do; the machines just know.

This paper suggests our brains work like the second factory.

What Did the Authors Discover?

The authors looked at how we judge the angle of lines (like the tilt of a picture frame). They found that our brains make predictable "mistakes" (biases).

• The Mistake: If a line is slightly tilted, we often think it's tilted more toward the vertical or horizontal (the "safe" angles we see most often).
• The Old Theory: Scientists thought this happened because our brain's "clue machine" was uneven (some angles were easier to see than others) AND because we had a separate "expectation" that pushed our judgment.
• The New Theory: The authors showed that you don't need two separate things. If you build the "clue machine" (the neurons in your eye) correctly, the "expectation" is automatically part of the machine's design.

The "Magic" Mechanism: Baking the Recipe into the Cake

The authors used a mathematical trick to show this. Imagine you are baking a cake.

• The Ingredients: The neurons in your brain.
• The Recipe: The "Prior" (what you expect to see).
• The Old View: You bake a plain cake, and then someone adds a special sauce (the expectation) on top.
• The New View: You mix the special sauce into the batter before baking. The cake is the sauce.

In the brain, this means the neurons are tuned so that their total activity naturally matches what we expect to see in the world.

• If vertical lines are common in our world, the brain has more neurons (or neurons that work harder) tuned to vertical lines.
• Because of this setup, when the brain tries to figure out "What did I just see?", it doesn't need to add a separate "expectation" step. The answer comes out naturally, automatically biased toward what is common.

The "Whitening" Effect: Why the Factory Looks Weird

The paper also looked at real data from cat brains (cats are great for studying vision).

• The Prediction: If the brain builds its factory to be super efficient, the total noise should be balanced out.
• The Surprise: When they looked at individual neurons, they were all working at the same steady pace (like a factory where every worker has the same job).
• The Twist: However, when they looked at groups of workers, they found that there were simply more workers assigned to the "common" angles (like vertical and horizontal) than the "rare" angles.

So, the brain doesn't make the workers work harder; it just hires more of them for the important jobs. This creates a "population bias" that looks exactly like a built-in expectation.

Why Does This Matter?

1. Efficiency: It saves energy. The brain doesn't need to store a separate "rule book" of expectations. The rules are written into the hardware itself.
2. Simpler Math: It explains why we make those specific "mistakes" (biases) without needing to assume the brain is doing complex, separate calculations. The bias is a natural side effect of an efficient system.
3. Real-World Proof: By re-analyzing old data from cat brains, the authors found that nature actually does it this way. The brain really does hire more neurons for common things, effectively "baking" the world's statistics into our vision.

The Takeaway

Think of your brain not as a computer that calculates expectations, but as a custom-built instrument. Just as a violin is shaped to make certain notes sound beautiful, your visual system is shaped by your life experiences to make certain sights "pop" out naturally. The bias isn't a bug; it's a feature of a highly efficient design.

Technical Summary

1. Problem Statement

The paper addresses a fundamental disconnect in visual neuroscience between Bayesian computational models and biological neural implementation.

• The Conflict: Standard Bayesian models treat "priors" (environmental expectations) and "neural resources" (encoding precision/Fisher information) as separable quantities. Recent work (Hahn & Wei, 2023) successfully modeled behavioral biases as the sum of two distinct forces: attraction to a prior and repulsion from peaks in neural resource allocation (Fisher information). However, this model often requires fitting a "flat" (uniform) explicit prior to match empirical data, attributing all biases to resource inhomogeneities.
• The Biological Question: Is it biologically necessary to store priors and likelihoods separately? The authors argue that efficient coding theories suggest the brain should internalize environmental statistics directly into the neural architecture (tuning curves, gain, density) to save metabolic energy. If priors are embedded in the encoder, the decoder does not need an explicit prior term.
• The Core Hypothesis: Behavioral biases (specifically the transition from repulsion to attraction) can be explained entirely by an efficient population code where the prior is embedded in the encoding properties, allowing for Bayes-optimal inference using only a uniform explicit prior during decoding.

2. Methodology

The authors employed a combination of theoretical modeling, mathematical derivation, and re-analysis of existing neurophysiological data.

• Theoretical Framework:

  • They formalized the relationship between the encoder and the posterior distribution using Bayes' theorem: $\log p(s|r) = \log p(r|s) + \log p(s) - \log p(r)$.
  • They proposed a constraint for an efficient sensory encoder: The sum of the population's tuning curves ($\gamma(s)$) must be proportional to the log-prior of the stimulus domain ($\log p(s)$).
  • Mechanism: They utilized a homeostatic scaling rule (normalization) where each neuron maintains a constant total spike budget under uniform input. This scales the gain of tuning curves to match the prior density without altering the Fisher information profile relative to the prior.

• Simulation & Modeling:

  • Task: Visual orientation estimation (reporting the mean orientation of Gabor arrays).
  • Prior: A bimodal distribution ($p(s) \propto 2 - |\sin(s)|$), reflecting the "oblique effect" (more sensitivity to cardinal orientations).
  • Encoder Design: Created a population of independent Poisson neurons. The tuning curves were shaped by the prior, and their gains were normalized to satisfy the condition $\gamma(s) \propto \log p(s)$.
  • Decoder: A Maximum A Posteriori (MAP) estimator using a flat (uniform) explicit prior.
  • Comparison: The model's output was compared against empirical human behavioral data and the original fits from Hahn & Wei.

• Neural Data Re-analysis:

  • Re-analyzed a meta-analysis of orientation-tuned neurons in cat V1 (Tolhurst et al., 1981/1982 data).
  • Analyzed three metrics: (a) Mean spike rate per neuron vs. preferred orientation, (b) Number of cells per orientation, and (c) Total population spike rate per orientation sub-population.

3. Key Contributions

1. Unification of Encoding and Decoding: The paper demonstrates that "priors" and "neural resources" are not separable in efficient biological codes. By embedding the prior into the tuning curve gain and density, the system achieves Bayes-optimal inference without a separate prior term in the decoder.
2. Mathematical Reduction of Bias: The authors show that under this efficient coding scheme, the standard two-term bias equation (Attraction + Repulsion) reduces to a single-term equation dependent solely on the Fisher information profile of the embedded encoder:
   $$B(s) = \beta \left( \frac{1}{J(s)} \right)'$$
   This implies that the transition from repulsive to attractive biases is driven by the interaction of stimulus noise with the embedded prior, not by tuning a loss-function exponent in the decoder.
3. Metabolic Efficiency Argument: The model provides a metabolically efficient explanation for "flat" behavioral priors. A flat prior in the decoder does not mean the brain lacks expectations; rather, it means the expectations have been successfully pre-computed and stored in the sensory encoder's architecture.

4. Results

• Behavioral Fit: The proposed model (solid lines in Figure 1f) successfully reproduced the empirical human data for orientation estimation. It captured the repulsion effect at low noise (driven by resource peaks) and the attraction effect at high noise (driven by the embedded prior), matching the performance of Hahn & Wei's more complex two-term model.
• Decoding Simplification: The simulation confirmed that when the sum of tuning curves is proportional to the log-prior, a decoder with a uniform prior produces the same optimal estimates as a decoder with an explicit, complex prior.
• Neural Predictions & V1 Data:
  • Prediction: An efficient decoder requires heterogeneous aggregate population activity (higher total spikes for orientations with higher prior probability), even if individual neurons are "whitened" (constant firing rates).
  • Observation: Re-analysis of cat V1 data showed that while individual neuron spike rates were relatively consistent (whitened), the number of neurons tuned to specific orientations varied significantly (anisotropy). Consequently, the total population activity was heterogeneous, peaking at orientations corresponding to the environmental prior.
  • Conclusion: The biological data supports the hypothesis that the brain implements the efficient decoding constraint via population density (number of cells) rather than individual gain modulation, effectively embedding the prior in the neural resource allocation.

5. Significance

• Reframing Perceptual Bias: The paper challenges the view that behavioral biases are merely errors or separate cognitive processes. Instead, it posits that systematic biases are a natural byproduct of an efficiently organized sensory code where the brain optimizes for metabolic cost and information transmission.
• Bridging Levels of Explanation: It provides a concrete mechanistic link between abstract Bayesian inference and concrete neural physiology (tuning curves, spike counts, and population density).
• Implications for Neuroscience: The findings suggest that "flat" priors in behavioral experiments should not be interpreted as a lack of expectation, but potentially as a sign of a highly efficient encoding strategy. This shifts the focus of research from "how the brain applies a prior" to "how the brain's architecture pre-computes the prior."
• Future Directions: The study offers testable predictions for neurophysiology, specifically regarding the relationship between population density, tuning curve gain, and aggregate spike rates in early sensory cortices.