Uncertainty estimation with prediction-error circuits

Loreen Hertäg, Katharina A. Wilmes, Claudia ClopathNature Communications 2025 · DOI

TL;DR

This paper proposes a hierarchical neural network model demonstrating how the brain uses prediction-error (PE) neurons to estimate the uncertainty of both sensory inputs and internal expectations. By leveraging the distinct response patterns of positive and negative PE neurons, the model explains how neural circuits dynamically weight information and provides a biological mechanism for the "contraction bias" observed in human perception.

What Problem Does This Paper Address?

To navigate a complex world, the brain must constantly integrate what it sees (sensory input) with what it expects (predictions). Ideally, it should trust the source that is more reliable. For example, in a thick fog, you should rely more on your memory of a familiar road than on your obscured vision.

While Bayesian theories suggest the brain performs this "uncertainty-weighted integration" optimally, the actual biological implementation has remained a mystery. Specifically, how do simple neural circuits calculate the variance (a proxy for uncertainty) of a signal in real-time? This paper tackles the gap between high-level cognitive theory and low-level circuit architecture.

Methods

The authors developed a rate-based mean-field network model featuring a hierarchy of prediction-error circuits.

  • PE Neurons: The core of the model relies on two types of excitatory neurons: negative PE (nPE) neurons, which fire when sensory input is weaker than predicted, and positive PE (pPE) neurons, which fire when input is stronger.
  • Hierarchical Architecture: The model uses a "lower" circuit to estimate the mean and variance of sensory stimuli and a "higher" circuit to do the same for the predictions themselves.
  • Interneuron Diversity: The circuit incorporates three types of inhibitory interneurons—PV, SOM, and VIP—to maintain the delicate excitation-inhibition (E/I) balance required for PE neurons to function.
  • Mathematical Innovation: The authors demonstrate that the squared sum of nPE and pPE activity can mathematically represent the variance of the input signal, which is then integrated by downstream "V neurons."

Datasets

As a computational modeling study, the paper does not use traditional large-scale datasets. Instead, it uses:

  • Simulated Stimuli: Sequences of step-wise constant inputs drawn from uniform and normal distributions to test the network's ability to track changing means and variances.
  • Literature Comparison: The model's results are validated against known experimental data regarding rodent primary sensory areas and human behavioral studies on perceptual biases.

Key Insights and Ideas

  • PE Neurons as Variance Calculators: The most striking insight is that the same neurons used to signal "something is wrong" (PE neurons) also provide the raw data necessary to calculate "how much can I trust this signal?" (uncertainty).
  • The Role of Inhibition: Different interneurons act as "tuning knobs" for uncertainty. For instance, activating PV neurons decreases the estimated variance, effectively biasing the brain to trust its predictions more. Activating VIP neurons pushes the system toward a 50/50 split between sensory input and prediction.
  • Contraction Bias Explained: The model naturally reproduces the "contraction bias"—the tendency to overestimate small values and underestimate large ones. The authors show this isn't necessarily a flaw, but a byproduct of how circuits integrate history and current input over time.
  • Neuromodulation: The model suggests that neuromodulators like Acetylcholine (ACh) or Noradrenaline (NA) might control our "trust" in sensory data by targeting specific interneuron populations.

Key Results

  • Accuracy: The network correctly estimated the mean and variance across a wide range of input statistics, with normalized absolute errors often below 5-10%.
  • Dynamic Weighting: The model successfully shifted its "sensory weight" ($\alpha$): when sensory noise increased, $\alpha$ dropped toward 0 (relying on predictions); when the environment became volatile, $\alpha$ increased toward 1 (relying on new sensory data).
  • Interneuron Perturbation: Simulations showed that perturbing specific interneurons could systematically bias perception, offering a potential framework for understanding the "pathological weighting" of inputs seen in conditions like autism or schizophrenia.

Limitations and Future Work

  • Change-Point Lag: The model is currently suboptimal immediately after a sudden environmental change (a "change point"). It takes time for the "memory neurons" to update, leading to temporary inaccuracies. Future work could integrate explicit change-point detection mechanisms.
  • Biological Validation: While the model predicts the existence of "V neurons" (variance-encoding neurons), these have not yet been definitively identified in biological circuits. Experimentalists could use the model's predictions to hunt for these neurons using optogenetics and calcium imaging.
  • Complexity: The current model is a simplified mean-field representation. Scaling this to more complex, high-dimensional sensory tasks (like visual object recognition) remains a significant future challenge.

Summary updated

Abstract

Neural circuits continuously integrate noisy sensory stimuli with predictions that often do not perfectly match, requiring the brain to combine these conflicting feedforward and feedback inputs according to their uncertainties. However, how the brain tracks both stimulus and prediction uncertainty remains unclear. Here, we show that a hierarchical prediction-error network can estimate both the sensory and prediction uncertainty with positive and negative prediction-error neurons. Consistent with prior hypotheses, we demonstrate that neural circuits rely more on predictions when sensory inputs are noisy and the environment is stable. By perturbing inhibitory interneurons within the prediction-error circuit, we reveal their role in uncertainty estimation and input weighting. Finally, we link our model to biased perception, showing how stimulus and prediction uncertainty contribute to the contraction bias.
BibTeX
@inproceedings{hertag2025uncertainty,
  author = {Loreen Hertäg and Katharina A. Wilmes and Claudia Clopath},
  title = {{Uncertainty estimation with prediction-error circuits}},
  booktitle = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-58311-6},
}
Figure