University of Rochester Study Shows Visual Learning Increases Information Redundancy, Challenging Classic Coding-Efficiency Views

Forum topic · ✨步子哥 · 2026-04-28

Summary

A research team at the University of Rochester, led by Shizhao Liu, has challenged the long-standing "code de-duplication" view in neuroscience, which holds that learning improves sensory encoding by reducing redundancy between neural responses. Training rhesus monkeys on visual discrimination tasks while recording V4 cortical activity with 96-channel microelectrode arrays, the researchers measured redundancy via Fisher Information, defined as the difference between shuffled and real population information. Contrary to classical predictions, learning significantly increased information redundancy, and redundancy strongly correlated with behavioral improvement across four learning stages. Redundancy increased only when monkeys actively performed the task, not during passive viewing, and was accompanied by enhanced per-neuron information content rather than information loss. The authors argue these findings support a Bayesian generative inference framework, in which feedback connections from higher decision areas share prior beliefs across V4 neurons, producing correlated activity. The results suggest a paradigm shift from discriminative, feedforward models toward generative accounts of perception, with implications for understanding learning disorders such as ADHD and for designing AI systems that incorporate generative feedback loops alongside feedforward discriminative architectures.

University of Rochester Team Reveals Information Redundancy Mechanism in Visual Learning, Challenging the Classic Coding-Subtraction View

The Classic View: Learning Reduces Redundancy to Improve Coding Efficiency

For decades, neuroscience has held a classic "coding-subtraction" view of how the brain optimizes sensory information for decision-making. According to this view, learning improves coding efficiency by reducing redundancy between neuronal responses. Early models assumed each neuron's information capacity is fixed (constrained by metabolic costs), so redundancy—repeated representation of information across neurons—was treated as inefficiency. Learning was thus seen as an optimization process that removes redundancy, including suppressing noise correlations, so downstream decision circuits can read out information more cleanly. This thinking also shaped machine vision, where many models treat visual processing as a feedforward discriminative process that transforms pixels into increasingly class-friendly representations.

The New Study: Learning Increases Redundancy, Supporting a Bayesian Generative Inference Framework

A University of Rochester team led by Shizhao Liu challenged this view. They trained rhesus monkeys on visual discrimination tasks and recorded neural responses in visual cortical area V4 using 96-channel microelectrode arrays. The surprising result: learning significantly increased information redundancy between neurons rather than decreasing it—consistent with Bayesian inference frameworks and contrary to coding-subtraction predictions.

1. Quantifying Redundancy and Tracking Learning

The researchers quantified population information using Fisher Information and defined redundancy as:

I_redundancy = I_shuffle - I_real

where I_shuffle is the information after shuffling trial order to remove correlations and I_real is the actual population information. Positive redundancy indicates correlations that reduce total information. They tracked V4 redundancy dynamically over weeks of training.

2. Strong Positive Correlation Between Redundancy and Performance

As monkeys learned, V4 information redundancy rose significantly—and redundancy strongly correlated with task performance: better performance meant higher redundancy. This held across four learning stages (two monkeys, two tasks: "cardinal" and "oblique" direction discrimination). In three stages the monkeys approached ideal-observer decision strategies; in one they remained suboptimal, yet redundancy rose in all cases.

3. Active Task Engagement Is Necessary

Redundancy increased only when monkeys actively performed the task. Passive viewing of the same stimuli produced no rise in redundancy, indicating the effect is tied to the brain's active inference process, not mere repeated exposure.

4. Redundancy Did Not Degrade Information—It Enhanced It

Despite greater shared information, total information was not impaired. Instead, redundancy increases were accompanied by higher information content in individual neurons. This supports the Bayesian account: redundancy is not mere duplication but the sharing of prior information that strengthens each neuron's posterior belief about the stimulus.

Generative Inference: Feedback-Driven Redundancy Increases

The findings support the idea that visual processing is a generative inference ("analysis-by-synthesis") process: rather than passively extracting features bottom-up, the brain actively uses internal generative models, via top-down processing, to interpret sensory input. In this framework, learning refines an internal model that generates expectations about sensory input.

Crucially, feedback connections transmit prior beliefs. During learning, higher decision areas feed task-relevant priors to sensory cortex, where they combine with sensory evidence to form posterior estimates in V4. Because priors are shared across neurons, their activity becomes correlated—i.e., redundancy increases. This feedback-driven sharing makes neurons "work together," in sharp contrast to the classic expectation that learning makes neurons more independent and efficient.

Implications for Perception Theory and AI

Theoretical: A Paradigm Shift from Coding Subtraction to Generative Inference

The results show that in the visual system, increased redundancy accompanies successful learning, challenging the classic view and pushing perception theory toward generative inference. Perception becomes an active process of constructing explanations of the world, with top-down priors enriching sensory representations.

Applied: Learning Disorders and Artificial Intelligence

Learning disorders: If the brain optimizes learning by increasing redundancy, some disorders may involve failures of this mechanism. For example, in ADHD the brain may fail to use feedback to share priors, yielding insufficient redundancy during learning.
AI: Most modern AI systems (e.g., convolutional neural networks) are feedforward and lack the brain's recurrent generative feedback. Introducing generative feedback loops—top-down modulation of lower-level perception—could allow faster learning from limited data, greater robustness under uncertainty, and flexible task adaptation, pointing toward AI that fuses discriminative and generative inference.

Conclusion

This study reveals a counterintuitive but crucial phenomenon: the brain increases information redundancy during visual learning. The finding strongly supports the generative inference framework, in which feedback distributes information across more neurons while enhancing individual neurons' information content—prompting a rethinking of perception as active world-model construction rather than passive feature extraction, and opening new directions in neuroscience and AI research.