Task learning increases information redundancy of neural responses in macaque visual cortex

How does the brain optimize sensory information for decision-making in new tasks? One hypothesis suggests learning reduces redundancy in neural representations to improve efficiency, while another, based on Bayesian inference, predicts learning increases redundancy by distributing information across neurons. We tested these hypotheses by tracking population responses in macaque cortical area V4 as monkeys learned visual discrimination tasks. We found strong support for the Bayesian predictions: task learning increased redundancy in neural responses over weeks of training and within single trials. This redundancy did not reduce information but instead increased the information carried by individual neurons. These insights suggest sensory processing in the brain reflects a generative rather than discriminative inference process.

💡 Research Summary

This paper investigates how visual discrimination learning reshapes neural population coding in macaque area V4, directly testing two competing theoretical frameworks. The classic “efficiency” view predicts that learning reduces redundancy—correlated variability that limits information—thereby making representations more compact for downstream read‑out. In contrast, Bayesian hierarchical generative inference predicts that learning increases redundancy because feedback conveys prior beliefs, causing neurons to share information about posterior estimates.

Two rhesus monkeys were trained on two orientation‑discrimination tasks: a cardinal task (0° vs 90°) and an oblique task (45° vs 135°). Stimuli contained dynamic noise, and task difficulty was manipulated via orientation coherence. Learning was divided into four epochs; a “learning index” (product of similarity to an ideal observer strategy and stimulus‑explained choice variance) quantified progress and correlated tightly with behavioral accuracy.

Neural activity was recorded with 96‑channel Utah arrays, yielding 35–83 units per session (single‑ and multi‑units). Direction tuning was evident, and ~40 % of units were significantly task‑responsive. The authors quantified information using linear Fisher information. To isolate the contribution of correlations, they computed I_shuffle (Fisher information after trial shuffling, which destroys inter‑neuron correlations) and I_real (the true population Fisher information). Redundancy was defined as I_redundancy = I_shuffle − I_real; a value near zero indicates independent responses, while positive values indicate information‑limiting correlations (i.e., redundancy).

Key findings:

1. Redundancy grows with learning. At epoch 1, I_redundancy was close to zero, matching prior reports of near‑independent variability. By later epochs, I_redundancy was significantly positive and strongly correlated with the learning index (Spearman ρ≈0.6–0.75, p < 10⁻⁴).

2. Task‑dependent effect. In the same recording sessions, passive‑viewing trials showed no systematic increase in I_redundancy, indicating that the rise is linked to active task engagement and likely driven by feedback from decision‑making areas.

3. Selectivity matters. Splitting the population by d′ revealed that high‑selectivity neurons exhibited a larger increase in redundancy, suggesting that task‑relevant cells are preferentially recruited into the shared posterior representation.

4. Within‑trial dynamics. When trials were divided into eight 200 ms bins, redundancy increased across the trial only in late‑learning sessions, consistent with the generative‑inference prediction that posterior beliefs accumulate over time and become shared among neurons.

5. Model validation. Simulations of a hierarchical Bayesian network reproduced the empirical pattern: both I_shuffle and I_real rose with learning, but I_shuffle grew faster, producing positive I_redundancy. The increase in I_shuffle reflected higher marginal Fisher information per neuron, confirming that each neuron carries more task‑relevant information after learning, which is then redistributed across the population.

6. Behavioral vs neural information. Converting psychometric performance to Fisher information (I_behav) showed that for the cardinal task I_behav exceeded I_real, implying that the animal exploits information from neurons not recorded. For the oblique task, I_behav was comparable or lower than I_real, suggesting less efficient use of the available neural code.

Overall, the data falsify the classic efficiency hypothesis and provide strong support for the Bayesian generative‑inference account: learning does not prune correlations to make the code sparser; instead, it creates structured, information‑sharing correlations that increase redundancy while simultaneously boosting the amount of information each neuron conveys. This suggests that sensory cortex implements a posterior belief representation, continuously reshaped by top‑down feedback, to support flexible decision‑making.