Efficient Brain Coding: Prior Attraction and Adapter Repulsion Unified

> Paper: Fast efficient coding and sensory adaptation in gain-adaptive recurrent networks > Authors: Arthur Prat-Carrabin, Maximilian V. Harl, Samuel J. Gershman > Published: Nature Communications, 2026-05-15 > DOI: 10.1038/s41467-026-73032-0

1. A Century-Old Contradiction: Attraction or Repulsion?

Neuroscience has long faced a puzzle. In theory, if the brain follows the efficient coding principle—maximizing information transmission—neuronal tuning curves should shift *toward* the most frequent stimuli in the environment. This is called prior attraction.

Experiments show the opposite. When animals are repeatedly shown the same visual stimulus (e.g., a grating of a particular orientation), neurons in primary visual cortex (V1) whose tuning curves lie near that stimulus shift *away* from the high-frequency stimulus—adapter repulsion. Repulsion is also accompanied by secondary effects: tuning curves near the adapter broaden, while distant ones narrow. Such rapid deformation, occurring within tens of milliseconds, does not match the timescale of traditional synaptic plasticity.

Which theory is right? This paper's answer: both—and they are fundamentally the same thing.

2. Core Finding: One Principle, Two Manifestations

The authors propose a gain-adaptive recurrent network. The key insight:

> Attraction and repulsion are not two separate neural mechanisms. They are the optimal solutions of a single efficient-coding objective—minimizing decoding error plus firing cost—under different statistical distributions.

2.1 Wide prior → prior attraction

When the stimulus distribution is broad, the optimized gain profile is inverted-U shaped: gain is largest where prior probability is highest. This gain profile spontaneously pulls effective tuning curves toward the prior's center; the narrower the prior, the stronger the clustering. This is the classic prior attraction predicted by efficient coding.

2.2 Sharp prior → adapter repulsion

Applying the same objective to a very sharp prior (mimicking a repeatedly presented adapter) yields a different optimum: an M-shaped, bimodal gain profile, with a gain dip at the adapter's center and two high-gain peaks on either side. These off-center peaks pull nearby tuning curves apart, naturally producing adapter repulsion.

2.3 Unified framework

| Condition | Gain profile | Tuning curve change | Phenomenon | |---|---|---|---| | Wide prior | Inverted-U (unimodal) | Contract toward center | Prior attraction | | Sharp prior | M-shaped (bimodal) | Shift away from adapter | Adapter repulsion |

Both are solutions to the same optimization problem; only the prior's shape differs.

3. Mechanism: Gain, Not Synapses

3.1 Problems with traditional accounts

Existing adapter-repulsion models rely on changes in synaptic weights—modifying all O(N²) connections. But synaptic plasticity is too slow (seconds to minutes), hard to reverse quickly, and requires implausibly coordinated changes across many synapses.

3.2 Advantages of gain modulation

In this model, the recurrent weight matrix is entirely fixed; only each neuron's feedforward gain—a scalar (O(N) parameters)—changes. Like a volume knob: no rewiring, just amplification. Crucially, in a recurrent network, gain changes propagate through recurrent connections, producing globally coordinated adaptation.

3.3 Mathematical mechanism

In a recurrent network, a neuron's effective tuning center is a weighted average of all neurons' feedforward positions. A non-uniform gain distribution means high-gain neurons exert "pull" on their neighbors, and the direction of tuning shifts is determined by the spatial gradient of gains. This is the mathematical basis for fast adaptation without changing synaptic weights.

4. Behavioral Evidence: Humans Recalibrate Within a Single Trial

To test whether sensory systems adjust encoding to priors on very short timescales, the authors ran a human numerosity estimation experiment. On each trial, subjects were cued with a prior distribution (narrow, medium, or wide) and then asked to estimate the number of briefly flashed dots.

Result: the variance of estimates scaled linearly with the width of the cued prior. Even when the prior changed randomly from trial to trial (seconds), humans rapidly reallocated internal representational resources—strong behavioral evidence that the brain can perform efficient coding in real time.

5. Relation to Existing Models

5.1 Comparison with Duong et al. (2023)

Duong et al.'s 2023 preprint (arXiv:2305.19869) proposed a similar gain-adaptive recurrent network, but focused on adapter repulsion and decorrelation, without explicitly contrasting sharp vs. wide priors and without behavioral evidence. This paper extends it by unifying attraction and repulsion, predicting adaptation patterns under different priors, and validating predictions with human behavior.

5.2 Contrast with synaptic plasticity models

| Feature | Synaptic plasticity models | Gain-adaptive model | |---|---|---| | Parameter count | O(N²) | O(N) | | Adaptation speed | Slow (seconds–minutes) | Fast (milliseconds) | | Reversibility | Poor | Good | | Biological plausibility | Requires widespread coordinated changes | Local gain adjustment suffices |

6. Theoretical Significance

The work breaks a long-standing conceptual split: prior attraction was seen as a product of efficient coding, while adapter repulsion was attributed to fatigue or inhibition. Now both appear as manifestations of one principle under different conditions—reminiscent of wave-particle duality in physics.

6.1 For neural engineering

Gain adaptation suggests a route to fast-adapting neural networks: no costly weight updates, just tuning scalar gain parameters for millisecond-scale network reconfiguration.

6.2 For understanding the brain

The study supports a dynamic computation view: the brain is not a fixed machine but a flexible system that reshapes its representational space in real time—not by rewiring (synapses), but by adjusting volume (gain).

7. Limitations and Open Questions

• Downstream decoding: If tuning curves change on millisecond timescales, how do downstream areas know the current encoding scheme? This "meta-representation" problem is beyond the paper's scope.
• Biological implementation of gain: The paper proves sufficiency computationally, but the biological substrate (neuromodulators like acetylcholine or dopamine, intrinsic excitability changes, or local network feedback) remains unclear.
• Rate vs. spiking models: The model is a linear firing-rate model; whether gain adaptation works the same way in spiking networks is an important extension.

8. Conclusion: An Elegant Unification

The paper's elegance lies in a minimal model—fixed recurrent connections plus variable neuronal gains—explaining two seemingly contradictory phenomena. The deepest insights in neuroscience often come from finding the right level of abstraction: identifying unified computational principles rather than simulating every ion channel. The brain achieves rapid adaptation with a volume knob rather than rewiring—a fitting metaphor for the beauty of biological intelligence.

---

Reference: Prat-Carrabin, A., Harl, M.V. & Gershman, S.J. Fast efficient coding and sensory adaptation in gain-adaptive recurrent networks. *Nat Commun* (2026). https://doi.org/10.1038/s41467-026-73032-0