Hierarchy in neuronal representations of multiple tasks in prefrontal cortex

By training macaque monkeys on multiple tasks and recording prefrontal cortex activity, researchers discovered that flexible multitasking is supported by a three-level hierarchical neural geometry where location codes are nested within subtask spaces that share a generalized ring-like structure, enabling robust generalization across both spatial locations and cognitive demands.

The Gist

Imagine your brain is a massive, bustling airport control tower. Every day, it has to manage thousands of different flights (tasks): landing a plane, refueling it, guiding a passenger to the gate, and checking the weather. The big mystery scientists have faced for years is: How does the control tower keep all these different jobs organized without crashing into each other?

This paper, titled "Hierarchy in Neuronal Representations of Multiple Tasks in Prefrontal Cortex," acts like a detective story. The researchers went inside the "control tower" of a monkey's brain (specifically the Prefrontal Cortex, the CEO of the brain) to see how it handles doing four different things at once.

Here is the story of what they found, explained simply:

1. The Experiment: The Monkey's "Swiss Army Knife" Test

The researchers trained two monkeys to play four very different video games on a screen.

• Game 1 (Discrimination): Touch a mole, ignore a snake.
• Game 2 (Attention): Look at a red dot, then touch a target.
• Game 3 (Sequence): Touch spots in a specific order (1-3-5-7).
• Game 4 (Nested): Do a small task while waiting to do a big task.

Crucially, all these games used the exact same eight spots on the screen. The monkeys had to switch between these games instantly. The scientists used a high-tech "microscope" (two-photon imaging) to watch thousands of individual neurons fire as the monkeys played.

2. The Problem: The "Mixed-Up" Neurons

The scientists expected to find "specialist" neurons. They thought: "Maybe some neurons only care about the 'mole' and others only care about the 'snake'."

But that wasn't what they saw. Instead, most neurons were jacks-of-all-trades. A single neuron might fire when the monkey sees the mole, but it also fires when the monkey is playing the attention game or the sequence game. It was like a single air traffic controller trying to talk to a pilot, a fuel truck driver, and a baggage handler all at the same time.

If every neuron is doing everything, how does the brain keep the tasks separate? If they all mix together, the monkey should get confused and fail. But the monkeys were experts!

3. The Solution: The "Neural Geometry" (The Shape of Thought)

The researchers realized that to understand the brain, you can't just look at one neuron; you have to look at the shape the whole group of neurons makes together.

They discovered a beautiful, hidden structure called a Hierarchical Neural Geometry. Imagine a set of Russian nesting dolls, or a tree with three distinct levels:

• Level 1: The Ring (The "Where")
  Inside every task, the neurons form a perfect circle (ring) representing the eight locations on the screen. Whether the monkey is playing the "Mole Game" or the "Snake Game," the brain draws the same circle to represent "Left," "Right," "Up," and "Down."

  • Analogy: Think of this as the map. The map of the city is the same whether you are driving a taxi, riding a bike, or walking. The "location" code is shared and consistent.

• Level 2: The Subtask Spaces (The "What")
  Now, imagine that circle is floating inside a specific room.

  • The "Mole Game" circle is in the Red Room.
  • The "Sequence Game" circle is in the Blue Room.
    Even though the map (the ring) looks the same, the room it's in is different. The brain separates the tasks by shifting the "center" of the circle or rotating the room slightly.
  • Analogy: This is like having the same recipe (the ring) but cooking it in different kitchens (the subtasks). The ingredients are the same, but the context changes the outcome.

• Level 3: The Meta-Tasks (The "Why")
  Finally, the researchers found that these "rooms" are grouped into larger neighborhoods.

  • One neighborhood is for "Go/No-Go" decisions (Do I touch it or not?).
  • Another neighborhood is for "Reward" expectations (Will I get a juice treat?).
    The brain organizes the specific games into these broader categories.
  • Analogy: This is the manager's office. The manager doesn't care about the specific recipe; they care about the goal. "Are we making a dessert (Reward) or a salad (No-Reward)?"

4. Why This Matters: The "Best of Both Worlds"

This hierarchy solves the brain's biggest problem: Sharing vs. Separating.

• Sharing: By using the same "Ring" for locations, the brain doesn't have to relearn how to find "Left" or "Right" every time it switches games. It's efficient!
• Separating: By putting those rings in different "Rooms" and "Neighborhoods," the brain ensures that the "Mole Game" doesn't accidentally tell the monkey to touch the "Snake."

5. The "Error" Clue

The researchers also looked at when the monkeys made mistakes. They found that when a monkey messed up, the error happened at the top level first, and then trickled down.

• If the monkey thought, "Oh, I'm in the 'Go' neighborhood," but actually needed to be in the "No-Go" neighborhood, the whole chain of events went wrong.
• It's like a GPS giving you the wrong destination (Meta-task); even if the map (Location) is perfect, you'll end up in the wrong place.

The Big Takeaway

The brain isn't a messy pile of wires where everything gets mixed up. It is a highly organized, multi-level library.

1. The Books (Locations) are organized on shelves.
2. The Shelves (Subtasks) are organized into different sections.
3. The Sections (Meta-tasks) are organized by the type of story (Goal).

This structure allows us (and monkeys) to be incredibly flexible. We can switch from driving a car to playing chess to having a conversation without our brain crashing, because we have a hierarchical map that keeps our "where," "what," and "why" perfectly organized.

Technical Summary

1. Problem Statement

The ability to perform multiple tasks flexibly is a hallmark of general intelligence. While biological brains achieve this with remarkable adaptability, the underlying neural mechanisms remain unclear. Specifically, it is unknown how a single population of neurons in the Lateral Prefrontal Cortex (LPFC) can simultaneously achieve two seemingly contradictory goals:

1. Representational Sharing: Reusing neural codes to generalize across tasks and locations (efficiency).
2. Representational Separation: Maintaining distinct codes to prevent inter-task interference (accuracy).

Previous studies have identified "mixed-selectivity" neurons (responding to both rules and stimuli) and low-dimensional manifolds, but the geometric organization that balances sharing and separation across a complex multitask environment has not been fully elucidated.

2. Methodology

The authors employed a combination of behavioral training, high-resolution neural recording, and advanced geometric analysis.

• Subjects & Paradigm: Two macaque monkeys (Macaca fascicularis) were trained on four distinct cognitive tasks sharing the same eight spatial locations but varying in cognitive demands:

  1. Discrimination Task (DT): Go/No-Go decisions based on visual stimuli (mole vs. owl/snake).
  2. Attention Task (AT): Reaching based on congruent or incongruent attention cues.
  3. Sequential Reach Task (SR): Reaching a fixed sequence (1-3-5-7) vs. catch trials.
  4. Nested Reach Task (NR): A dual-task paradigm involving a delay reach (outer) and a sequential reach (inner) to induce interference.
  • Total: These tasks comprised 14 distinct subtasks (e.g., specific go/no-go conditions, reward predictions, nested vs. standalone trials).

• Neural Recording:

  • Technique: Two-photon calcium imaging using GCaMP6f expressed via AAV vectors in the LPFC.
  • Scale: Recorded from thousands of neurons (Monkey 1: ~2,424 neurons; Monkey 2: ~937 neurons) across multiple fields of view (FOVs).
  • Data: Analyzed population activity during specific task epochs (330–1000 ms post-stimulus).

• Analytical Framework:

  • Linear Regression (GLM): Modeled neural responses as a function of spatial location (one-hot encoded) and task context.
  • Geometric Decomposition:
    • Subtask Spaces: Extracted low-dimensional manifolds (2D) for each subtask using Principal Component Analysis (PCA) on regression coefficients.
    • Offsets: Analyzed regression bias terms to characterize the "center" of each subtask space.
  • Generalization Tests: Used Support Vector Machines (SVM) to test cross-task and cross-location decoding (e.g., "Leave-one-task-out" and "Leave-one-location-out").
  • Anatomical Analysis: Calculated clustering indices to determine if neurons with similar functional properties were spatially clustered within the cortex.

3. Key Contributions

1. Identification of a Three-Level Hierarchical Geometry: The study proposes a novel hierarchical structure organizing multitask representations:
   • Level 1 (Location): Spatial locations are encoded within low-dimensional subtask spaces exhibiting a shared ring-like structure.
   • Level 2 (Subtask): Different subtasks are separated by distinct orthonormal bases and offsets within the high-dimensional neural state space.
   • Level 3 (Meta-Task): Subtasks are grouped into broader meta-tasks (defined by "Go/No-Go" and "Reward/No-Reward" features).
2. Resolution of the Sharing-Separation Paradox: The paper demonstrates that sharing and separation are not mutually exclusive but are organized hierarchically. Shared spatial codes exist within subtask spaces, while separation is achieved via the offsets and bases of these spaces.
3. Meta-Task Dominance: The study identifies that meta-task features (abstract rules like motor planning and reward expectation), rather than specific task identities, primarily drive the separation between subtask spaces.
4. Anatomical-Functional Correlation: The authors provide evidence that the degree of anatomical clustering decreases as the level of abstraction increases (Location > Subtask > Meta-Task).

4. Key Results

• Single Neuron Limitations: Single neurons exhibited mixed selectivity (76.1% of neurons), encoding both location and task context. No single neuron type could fully explain task separation, necessitating a population-level geometric analysis.
• Shared Spatial Manifold:
  • Across 14 subtasks, the neural representations of the 8 spatial locations formed a consistent ring-like geometry.
  • A decoder trained on spatial locations in 4 subtasks generalized to the remaining 10 subtasks with high accuracy, confirming a shared neural manifold for spatial coding.
  • This sharing was robust to eye movement patterns (verified via an Intercept Reach task control).
• Task Separation via Meta-Tasks:
  • Subtask spaces were not orthogonal but showed varying degrees of separation.
  • Meta-task features (Go/No-Go and Reward/No-Reward) explained the majority of the variance in the distance between subtask spaces (CPD ~46% and ~36%, respectively).
  • In contrast, specific "task sets" (e.g., DT vs. AT) explained almost none of the distance variance, indicating that the brain organizes tasks by abstract rules rather than specific task identities.
• Hierarchical Generalization:
  • Cross-Task: Classifiers trained to distinguish Go/No-Go or Reward/No-Reward meta-tasks generalized perfectly across different tasks.
  • Cross-Location: Classifiers trained to distinguish subtasks generalized across different spatial locations.
• Error Behavior Explanation:
  • Monkey errors were correlated across hierarchical levels. When a monkey made an error at the subtask level (e.g., choosing "Go" instead of "No-Go"), the neural trajectory also deviated at the meta-task level.
  • The distance of error trials to decision boundaries at the subtask level strongly correlated with the distance at the meta-task level ($r \approx 0.87-0.90$), suggesting errors propagate through the hierarchy.
• Anatomical Organization:
  • Neurons with similar location preferences showed significant anatomical clustering within ~150 $\mu$m.
  • Clustering was weaker for subtask-level offsets and absent for meta-task-level offsets, suggesting a decreasing trend in anatomical specificity as abstraction increases.

5. Significance

• Mechanism of Cognitive Flexibility: The paper provides a concrete geometric mechanism for how the brain achieves "general intelligence." It suggests that the brain does not need separate circuits for every task; instead, it uses a hierarchical tree structure where low-level sensory codes are shared, and high-level abstract rules provide the necessary separation.
• Bridging Single-Cell and Population Views: It unifies the concepts of "mixed-selectivity neurons" and "low-dimensional manifolds," showing how mixed selectivity at the single-neuron level gives rise to structured, factorized representations at the population level.
• Implications for AI: The findings offer a blueprint for artificial intelligence systems. Current AI models often struggle with catastrophic forgetting or require massive retraining for new tasks. This study suggests that implementing hierarchical, factorized representations (separating sensory inputs from abstract rules via offsets and bases) could enable AI to learn new tasks rapidly with minimal interference, mimicking biological generalization.
• Clinical Relevance: Understanding the hierarchical breakdown of multitasking could inform research into cognitive disorders (e.g., ADHD, schizophrenia) where the balance between task sharing and separation is disrupted.