Desegregation of neuronal predictive processing
Bin Wang1,2, Nicholas J Audette3, David M Schneider3
1Department of Physics, University of California San Diego, La Jolla, CA, USA.
Nature Communications
|March 14, 2026
Summary
Neural circuits build internal world-models using predictive processing. Contrary to specialized cells, predictions and errors are integrated in distributed networks, advancing our understanding of brain computation.
Area of Science:
- Neuroscience
- Computational Neuroscience
- Cognitive Science
Background:
- Neural circuits form internal world-models to guide behavior.
- Predictive processing framework suggests neural activity signals predictions and computes prediction-errors.
Purpose of the Study:
- Investigate the emergence of high-dimensional, multi-modal predictive representations in recurrent networks.
- Understand how the brain generates predictions for complex sensorimotor signals.
Main Methods:
- Utilized recurrent neural networks to model predictive processing.
- Employed a rich stimulus-set to violate learned expectations in animals.
- Analyzed the distribution of stimulus and prediction-error representations.
Main Results:
- Stimulus and prediction-error representations are desegregated in networks performing robust predictive processing, challenging theories of specialized cell-types.
- Predictive processing is optimal when excitation/inhibition balance is loose.
- Distinct functional roles of excitatory and inhibitory neurons were revealed.
Conclusions:
- Neural representations of internal models are highly distributed yet structured for flexible behavioral readout.
- Demonstrated that neural representations of internal models are computed by incorporating diverse computations into a unifying model.
