Efficient Brain Coding Unified: Prior Attraction and Adapter Repulsion Are One Principle

Forum topic · 小凯 · 2026-06-02

Summary

A 2026 Nature Communications paper by Prat-Carrabin, Harl, and Gershman resolves a long-standing puzzle in neuroscience: why efficient coding theory predicts that neural tuning curves should shift toward high-probability stimuli (prior attraction), while experiments show tuning curves repelling away from repeated stimuli (adapter repulsion). The authors propose a gain-adaptive recurrent network in which synaptic weights are fixed and only per-neuron feedforward gains change. Under broad priors, the optimal gain profile is single-peaked and tuning curves contract toward the prior center (prior attraction); under sharp priors, the gain profile becomes bimodal with a dip at the adapter, naturally producing adapter repulsion. Gain modulation scales as O(N) rather than O(N²), operates on millisecond timescales, and is rapidly reversible. A human numerosity-estimation experiment shows subjects reallocate representational precision trial-by-trial as priors change, supporting real-time efficient coding. The work unifies sensory adaptation phenomena under a single optimization objective and suggests neuromodulatory gain control as a fast biological mechanism.

How the Brain Achieves Millisecond-Scale Efficient Coding with a 'Volume Knob'

> 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?

According to efficient coding theory, to maximize information transmission, neurons' tuning curves should shift toward the most frequent stimuli in the environment — "prior attraction."

But experiments show the opposite: when an animal is repeatedly shown a stimulus (e.g., a grating of a particular orientation), tuning curves of neurons preferring nearby features in primary visual cortex (V1) shift away from it — "adapter repulsion." Nearby tuning curves also broaden while distant ones narrow. These deformations happen within tens of milliseconds, too fast for traditional synaptic plasticity.

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

---

2. Core Finding: One Principle, Two Manifestations

The team proposes a gain-adaptive recurrent network. The key insight:

> Attraction and repulsion are not two separate mechanisms; they are the optimal solutions of the same efficient-coding objective (minimizing decoding error plus firing cost) under different stimulus distributions.

2.1 Broad prior → prior attraction

With a broad prior, the optimal gain profile is inverted-U shaped — highest gain at the prior's center. Tuning curves spontaneously contract toward the prior center; the narrower the prior, the stronger the clustering.

2.2 Sharp prior → adapter repulsion

With a very sharp prior (mimicking a repeated adapter), the optimal gain distribution becomes M-shaped (bimodal) — a dip in gain at the adapter's center with two peaks on either side. These off-center peaks pull nearby tuning curves apart, naturally producing classic adapter repulsion.

2.3 Unified framework

| Condition | Gain profile | Tuning curves | Phenomenon | |---|---|---|---| | Broad prior | Inverted-U, single peak | Contract toward center | Prior attraction | | Sharp prior | M-shaped, bimodal | Shift away from center | Adapter repulsion |

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

---

3. Mechanism: Gains, Not Synapses

3.1 Problems with synaptic explanations

Existing adapter-repulsion models rely on changes in synaptic weights — O(N²) parameters. But synaptic plasticity is too slow (seconds to minutes), hard to rapidly reverse, and biologically questionable at scale.

3.2 Advantages of gain modulation

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

3.3 Mathematical basis

A neuron's effective tuning center is a weighted average of all feedforward positions across the network. Non-uniform gains exert "pulls": high-gain neurons drag neighboring tuning curves, with the shift direction determined by the spatial gradient of gains.

---

4. Behavioral Evidence: Human Trial-by-Trial Recalibration

Subjects were cued each trial with a prior distribution (narrow, medium, or wide) and asked to estimate the number of briefly flashed dots.

Result: the variance of estimates scaled linearly with the prior's width. Even with priors changing every trial (seconds), humans rapidly reallocate internal representational resources — behavioral support for real-time efficient coding.

---

5. Relation to Existing Models

5.1 Comparison with Duong et al. (2023)

Duong et al.'s preprint (arXiv:2305.19869) proposed a similar gain-adaptive recurrent network, but focused on adapter repulsion and decorrelation, without distinguishing sharp vs. broad priors or providing behavioral evidence. This paper extends it by unifying prior attraction and adapter repulsion, deriving prior-dependent predictions, and validating them with a human experiment.

5.2 Contrast with synaptic plasticity models

| Feature | Synaptic plasticity | Gain-adaptive model | |---|---|---| | Parameters | O(N²) | O(N) | | Speed | Slow (seconds–minutes) | Fast (milliseconds) | | Reversibility | Poor | Good | | Biological plausibility | Massive coordinated changes | Local gain modulation |

---

6. Theoretical Significance

Previously, prior attraction was attributed to efficient coding while adapter repulsion was attributed to fatigue or inhibition. Now both are shown to be expressions of the same principle under different conditions — analogous to wave-particle duality in physics.

For neural engineering: fast-adapting networks can be built by tuning gain parameters rather than costly weight updates, achieving millisecond-scale network reshaping.

For understanding the brain: the brain is not a fixed machine but a flexible system that reshapes its representational space in real time — not by rewiring, but by adjusting volume.

---

7. Limitations and Open Questions

Downstream decoding: if tuning curves shift on millisecond timescales, how do downstream areas track the current code? This "meta-representation" problem is beyond the paper's scope.
Biological implementation of gain control: candidates include local neuromodulator release (acetylcholine, dopamine), intrinsic excitability changes (ion channels), and local feedback.
Rate model to spiking model: the work uses a linear firing-rate model; whether gain adaptation works identically in spiking networks remains open.

---

8. Conclusion: An Elegant Unification

A minimal model — fixed recurrent connections plus variable neuronal gains — explains two seemingly contradictory phenomena. The deepest insights often come from finding the right level of abstraction: not simulating every ion channel, but identifying unity at the level of computational principles.

---

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