Task learning increases information redundancy of neural responses in macaque visual cortex
Monkey brain research reveals task learning makes neurons more redundant, not less—opposite of current AI assumptions.
A team led by Shizhao Liu, Anton Pletenev, Ralf M. Haefner, and Adam C. Snyder published groundbreaking research in Science (2026) that fundamentally challenges how we understand learning in both biological and artificial systems. By tracking population responses in macaque cortical area V4 over weeks of visual discrimination training, they discovered that task learning doesn't optimize by reducing redundancy—instead, it increases redundancy by distributing information across more neurons. This finding provides strong experimental support for Bayesian inference models of brain function over traditional efficiency models.
The research revealed that this increased redundancy doesn't dilute information but actually enhances it, with individual neurons carrying more task-relevant information as learning progresses. The team observed these effects both across weeks of training and within single trials, suggesting the brain employs a generative inference process rather than a purely discriminative one. This has profound implications for artificial intelligence, where current architectures often prioritize sparsity and efficiency—directly opposing what appears to be the brain's actual strategy for robust learning.
For AI researchers, these findings suggest that deliberately building redundancy and distributed representations into neural networks might improve their robustness and learning capabilities. The study's methodology—combining long-term neural recording with information theory analysis—also provides a new framework for evaluating how artificial systems learn compared to biological ones. As AI continues to draw inspiration from neuroscience, this research indicates we may need to reconsider which aspects of brain function we're trying to replicate.
- Learning increased neural response redundancy in macaque V4 cortex over weeks of visual task training
- Findings support Bayesian inference models (predicting redundancy increase) over efficiency models (predicting redundancy reduction)
- Individual neurons carried more information despite increased redundancy, suggesting generative rather than discriminative processing
Why It Matters
Challenges fundamental AI assumptions about efficiency and sparsity, potentially leading to more robust neural network architectures inspired by biological learning.