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

By tracking macaque V4 neural responses during visual discrimination learning, the study demonstrates that task learning increases information redundancy rather than reducing it, supporting a Bayesian generative inference model where redundancy enhances the information carried by individual neurons.

The Gist

Here is an explanation of the paper using simple language and creative analogies.

The Big Question: How Does the Brain Learn?

Imagine your brain is a massive factory that processes information from your eyes. For decades, scientists believed that when you learn a new skill (like telling the difference between a horizontal line and a vertical one), the factory gets more efficient. They thought the brain would "fire fewer workers" and "cut out the noise," making every single neuron work perfectly on its own to save energy.

The "Classic" View: Learning = Cutting redundancy. Like a manager telling workers, "Stop talking to each other and just do your own job perfectly."

The New Discovery: This paper suggests the opposite is true. When the monkeys in the study learned the task, their brain cells didn't become more independent; they started talking to each other more. They became a highly coordinated team where everyone shared the same information.

The New View: Learning = Increasing redundancy. Like a manager saying, "Everyone, make sure you all know the exact same plan so we can't fail."

---

The Experiment: The Monkey "Line Detectives"

The researchers trained two rhesus monkeys to play a video game.

• The Game: A screen would flash a noisy, static-filled image. Sometimes the noise was tilted slightly to the left (0° or 90°), and sometimes slightly to the right (45° or 135°).
• The Goal: The monkey had to look at the screen and guess the direction of the tilt by looking at a specific target.
• The Tool: The scientists implanted a tiny "microphone array" (96 channels) into the monkeys' visual cortex (area V4) to listen to the electrical chatter of hundreds of neurons at once.

They watched the monkeys practice for weeks. At first, the monkeys were bad at the game. Over time, they got perfect at it.

The Surprise Finding: The "Choir" Effect

The scientists measured something called Information Redundancy.

• Low Redundancy: Neuron A knows the answer, and Neuron B knows a different piece of the puzzle. They are independent.
• High Redundancy: Neuron A, B, C, and D all know the exact same answer. They are all shouting the same thing.

What they found:
As the monkeys got better at the game, their neurons started shouting the same answer in unison. The "redundancy" went up, not down.

The Analogy:
Imagine a group of people trying to guess the weather.

• Before Learning (Classic View): Person A looks at the sky, Person B looks at the birds, Person C looks at the barometer. They all have different, unique data. If one person is wrong, the group still has the answer.
• After Learning (This Paper's Finding): The group learns that the birds are the best indicator. So, everyone starts watching the birds. Now, if you ask Person A, B, or C, they all say, "The birds are flying low, so it's going to rain." They are all saying the exact same thing.

Why is this good?
You might think, "But if they all say the same thing, isn't that a waste? What if they are all wrong?"
The paper argues that this "waste" is actually a superpower. By having everyone share the same "belief" about the world, the brain creates a safety net. Even if the signal is weak or noisy, the fact that everyone agrees makes the signal incredibly strong and reliable for the decision-making part of the brain.

The "Generative" vs. "Discriminative" Brain

The paper compares two ways the brain might work:

1. The Discriminative Brain (The Old Idea): The brain is a filter. It takes the messy world and tries to strip away everything unnecessary to find the "truth." It tries to be a perfect, efficient decoder.
2. The Generative Brain (The New Idea): The brain is a storyteller. It has a "story" (a model) of how the world works. When it sees something, it doesn't just decode it; it asks, "Based on my story, what should I be seeing?"

The monkeys learned to use this "story." They started using feedback loops.

• The Mechanism: The decision-making part of the brain (the "Boss") sends a message back down to the visual neurons (the "Workers"). The Boss says, "I think it's a horizontal line." The visual neurons hear this and adjust their own signals to match that belief.
• The Result: The visual neurons and the decision neurons start vibrating in sync. This creates the "redundancy" the scientists saw.

The "Within-Trial" Magic

Here is the coolest part: This redundancy didn't just happen over weeks of training; it happened within a single second of the game.

• Early in the trial: The monkey sees the image. The neurons are a bit confused.
• Later in the trial: As the monkey accumulates evidence, the "Boss" starts forming a belief. It sends that belief back down. The neurons update their signals to match the belief.
• The Result: By the end of the 1.6-second trial, the neurons are all singing the same song, making the final decision rock-solid.

The Takeaway

For a long time, we thought the brain learned by becoming more efficient and less repetitive. This paper suggests that learning is actually about building a strong, shared consensus.

When you learn a new skill, your brain doesn't just get better at "hearing" the signal; it gets better at broadcasting that signal to every corner of the relevant network. It creates a chorus of neurons all singing the same note, ensuring that the decision is made with maximum confidence.

In short: The brain learns by turning a group of independent soloists into a synchronized choir.

Technical Summary

Here is a detailed technical summary of the paper "Task learning increases information redundancy of neural responses in macaque visual cortex."

1. Problem Statement

The study addresses a fundamental debate in systems neuroscience regarding how the brain optimizes sensory representations for decision-making during learning. Two competing theoretical frameworks offer contradictory predictions:

• The "Classic" Perspective (Efficiency/Discriminative): Posits that learning and attention optimize neural codes by reducing redundancy. In this view, information-limiting correlations (noise correlations that limit the total information a population can carry) are viewed as harmful noise that should be suppressed to improve decoding accuracy.
• The "Generative Inference" Perspective (Bayesian): Posits that the brain performs probabilistic inference by inverting a generative model. In this framework, sensory neurons represent posterior beliefs that integrate incoming sensory evidence with prior expectations. This integration requires sharing information across neurons, which theoretically predicts that learning a task should increase information redundancy (manifested as increased correlated variability) to facilitate the propagation of task-specific priors.

The authors sought to empirically distinguish between these frameworks by tracking how neural population statistics in the visual cortex change as monkeys learn visual discrimination tasks.

2. Methodology

Subjects and Experimental Design:

• Subjects: Two adult male rhesus macaques (Macaca mulatta).
• Tasks: The monkeys were trained on two coarse-orientation discrimination tasks over several weeks:
  1. Cardinal Task: Discriminate between $0^\circ$and$90^\circ$.
  2. Oblique Task: Discriminate between $45^\circ$and$135^\circ$.
• Stimuli: Dynamic white noise stimuli where the orientation energy was modulated to create varying levels of "coherence" (signal strength). Stimulus duration was 1.6 seconds.
• Behavioral Tracking: A "Learning Index" was defined as the product of the similarity between the animal's empirical strategy and an ideal observer, and the degree to which choices were explained by the stimulus. This allowed for tracking learning progress even as stimulus difficulty (coherence) varied.

Neural Recording:

• Area: Visual cortical area V4.
• Hardware: 96-channel Utah arrays chronically implanted.
• Data: Population activity recorded during both active task performance and passive viewing sessions.
• Analysis: The study focused on Linear Fisher Information ($I$) to quantify information content.
  • $I_{real}$: Actual Fisher information carried by the correlated population.
  • $I_{shuffle}$: Fisher information calculated after shuffling trials to destroy noise correlations (representing the sum of independent marginal information).
  • Redundancy ($I_{redundancy}$): Defined as $I_{shuffle} - I_{real}$. A positive value indicates that correlations are limiting the total information (redundancy).

Modeling:

• The authors utilized a Hierarchical Bayesian Inference model (based on Haefner et al., 2016) to simulate neural sampling. This model predicts that learning a task involves updating task-specific priors, which are communicated via feedback, leading to increased redundancy.

3. Key Contributions

1. Empirical Validation of Generative Inference: The study provides strong empirical evidence that task learning increases information redundancy in sensory cortex, directly contradicting the classic "efficiency" hypothesis which predicts a decrease in redundancy.
2. Mechanism of Redundancy: The authors demonstrate that this increase in redundancy does not result from a loss of total information. Instead, it is driven by an increase in the marginal information carried by individual neurons ($I_{shuffle}$ increases faster than $I_{real}$). This suggests that learning redistributes information, making individual neurons more informative but more correlated.
3. Temporal Dynamics: The study reveals that redundancy increases not only over weeks of training but also within a single trial (over hundreds of milliseconds) as the animal accumulates sensory evidence.
4. Task Dependence: The increase in redundancy is specific to active task engagement. Passive viewing of the same stimuli did not show the same consistent increase in redundancy, indicating the effect is driven by dynamic, state-dependent feedback rather than fixed structural changes.

4. Key Results

• Redundancy Increases with Learning: Across all four experimental epochs (two monkeys $\times$ two tasks), $I_{redundancy}$ was near zero at the start of training and increased significantly as the monkeys learned the tasks.
  • Statistical Significance: Strong positive correlations were found between the Learning Index and $I_{redundancy}$ (e.g., Monkey R Cardinal: $\rho = 0.74, p = 9.3 \times 10^{-5}$).
• Source of Redundancy: Analysis of $I_{real}$ vs. $I_{shuffle}$ revealed that the rise in redundancy was driven by a rapid increase in $I_{shuffle}$ (marginal information per neuron) rather than a decrease in $I_{real}$. This confirms that individual neurons became more informative, and this shared information created redundancy.
• Within-Trial Dynamics: In late-learning sessions, $I_{redundancy}$ increased significantly over the course of a single trial (divided into 200ms bins). This supports the Bayesian prediction that neurons integrate evidence over time, sharing prior expectations as the trial progresses.
• Active vs. Passive: The increase in redundancy was significantly stronger during active task performance compared to passive viewing in late-learning stages, ruling out simple adaptation or feedforward plasticity as the sole cause.
• Neural vs. Behavioral Performance: The study found that behavioral performance (psychometric thresholds) was closely matched by the neural population's decoding capability (neurometric thresholds), suggesting the brain effectively utilizes this redundant, correlated code.
• Task-Informative Neurons: The effect was strongest in the 50% of neurons with the highest stimulus sensitivity ($d'$), indicating that the most relevant neurons are the ones sharing the most information.

5. Significance and Implications

• Paradigm Shift: The findings challenge the prevailing view that the brain optimizes sensory processing by "untangling" variables and minimizing correlations. Instead, it suggests the brain optimizes by propagating beliefs through a network of highly correlated neurons.
• Role of Feedback: The results support the hypothesis that feedback connections (from decision-making areas like parietal or prefrontal cortex) play a crucial role in broadcasting task-specific priors to sensory areas, thereby increasing redundancy to ensure robust decoding.
• Robustness: The authors suggest that increased redundancy may facilitate robust, non-specific decoding and faster learning, as the shared information allows downstream areas to read out task-relevant variables from smaller subsets of neurons.
• Interpretation of Correlations: The study reinterprets "information-limiting correlations" not as noise to be eliminated, but as a necessary byproduct of a generative inference process where neurons share probabilistic beliefs about the stimulus.

In conclusion, this paper demonstrates that the brain's strategy for learning new perceptual tasks is to increase the redundancy of neural representations through the sharing of task-specific priors, aligning with a Bayesian generative inference framework rather than a classic discriminative efficiency framework.