University of Rochester Study Shows Visual Learning Increases Information Redundancy, Challenging Classic Coding-Subtraction View

Forum topic · ✨步子哥 · 2026-04-28

Summary

A University of Rochester team led by Shizhao Liu trained rhesus macaques on visual discrimination tasks while recording V4 neural activity with 96-channel microelectrode arrays. Contrary to the classic view that learning improves coding efficiency by reducing redundancy between neurons, the study found that learning significantly increased information redundancy. Redundancy was quantified as the difference between Fisher information after trial-shuffling (removing correlations) and real population information. Redundancy rose in parallel with behavioral performance across four learning stages in two monkeys, and only increased when animals actively performed tasks, not during passive viewing. Importantly, increased redundancy did not reduce total information; instead, individual neurons carried more stimulus information. The authors argue this pattern supports a Bayesian generative inference framework, in which feedback connections share prior beliefs across neurons, rather than traditional feedforward discriminative models. Findings suggest implications for learning disorders such as ADHD and for AI architecture design, pointing to the value of generative feedback loops in perceptual systems.

Background: The Classic Coding-Subtraction View

Neuroscience has long held that learning optimizes sensory coding by reducing redundancy—the repeated representation of information across neurons—to improve encoding efficiency and downstream decoding. Under this view, each neuron's information capacity is fixed by metabolic constraints, so redundancy is treated as inefficiency to be removed. Machine vision models similarly treat perception as a feedforward, *discriminative* process that transforms pixels into classification-friendly representations.

New Findings: Learning Increases Redundancy

A University of Rochester team (Shizhao Liu et al.) trained rhesus macaques on visual discrimination tasks and recorded V4 responses using 96-channel microelectrode arrays. Results challenged the classic view: learning significantly increased inter-neuronal information redundancy, consistent with Bayesian generative inference rather than coding subtraction.

1. Quantifying Redundancy

Redundancy was defined via Fisher information:

I_redundancy = I_shuffle - I_real

where I_shuffle is information computed after shuffling trial order to remove correlations, and I_real is actual population information. Positive values indicate that correlations reduce total information (redundancy). Redundancy was tracked in V4 across weeks of training.

2. Redundancy Correlates Strongly with Learning Performance

As monkeys improved at the tasks, V4 redundancy increased significantly. Performance and redundancy were strongly positively correlated—better performance, higher redundancy. This held across four learning stages (two monkeys × two task variants: "principal direction" and "oblique direction" tasks), regardless of whether monkeys adopted near-ideal or suboptimal decision strategies.

3. Active Task Engagement Is Necessary

Redundancy increased only when monkeys actively performed the task. During passive viewing of identical stimuli, no redundancy increase was observed—indicating the effect depends on active inference and decision-making, not mere stimulus exposure.

4. Redundancy Enhances, Not Weakens, Information

Increased redundancy did not degrade total information. Instead, individual neurons carried more stimulus information, consistent with the Bayesian view that redundancy reflects *shared prior information* strengthening each neuron's posterior belief about the stimulus.

Generative Inference: Feedback-Driven Redundancy

The findings support generative inference ("analysis-by-synthesis"): rather than passively extracting features bottom-up, the brain actively uses internal generative models, via top-down feedback, to explain sensory input. Learning refines this internal model.

Feedback connections transmit prior beliefs from higher decision areas to sensory cortex; these priors combine with sensory evidence in V4 to form posterior estimates. Because priors are shared across neurons, activity correlations—and hence redundancy—increase. Discriminative models, which treat feedback as merely fine-tuning feedforward representations, predict redundancy should *decrease* with learning; the observed increase is only explained by generative inference.

Implications

1. Theoretical: A Paradigm Shift

The results challenge the long-standing "reduce redundancy" principle, supporting a shift toward generative inference: perception is the brain's active construction of explanations for the world, with top-down priors enriching sensory representations.

2. Applied: Learning Disorders and AI

Learning disorders: If the brain optimizes learning by increasing redundancy, conditions like ADHD may involve impaired use of feedback to share priors, yielding insufficient redundancy during learning. Future work could explore these neural redundancy abnormalities and interventions.
Artificial intelligence: Modern discriminative architectures (e.g., feedforward CNNs) lack the brain's recurrent generative feedback. Introducing generative feedback loops—top-down modulation of lower-level perception—could enable faster learning from limited data, greater robustness under uncertainty, and flexible task adaptation, pointing toward AI that fuses discriminative and generative inference.

Conclusion

The study reveals a counterintuitive but crucial phenomenon: the brain increases information redundancy during visual learning via feedback mechanisms that distribute priors across neurons, enhancing per-neuron information. This strongly supports generative inference frameworks over coding-subtraction accounts, reframing perception as active world-model construction rather than passive feature extraction.