Predictive abstract task representations support few-shot learning in the rat brain

Technical Summary: Predictive Abstract Task Representations Support Few-Shot Learning in the Rat Brain

Problem Statement
Few-shot learning—the ability to rapidly infer rules or strategies from sparse experience—is a hallmark of flexible cognition. While it is established that the orbitofrontal cortex (OFC) encodes abstract representations of task structure, it remains unknown whether these representations can adapt on the rapid timescale of few-shot learning (seconds to minutes) to reflect changing task contingencies. Previous studies have typically examined abstract representations over the slow timescale of days or sessions. This study addresses the gap between behavioral rapid adaptation and the neural mechanisms supporting it, specifically investigating if OFC representations of "roles" (the function of an action within a task structure) can be remapped as quickly as behavioral performance improves.

Methodology
The study utilized five female Sprague Dawley rats performing a sequence-learning task in the Rat Interactive Foraging Facility (RIFF), a circular arena with six interaction areas (IAs).

• Task Design: The task was modeled as a Markov Decision Process (MDP) with three states (State 1, State 2, and a Reward state) and six possible actions (pokes into specific IAs). In each "sequence block," two specific IAs were designated as correct ($a_1$ in State 1, $a_2$ in State 2) to transition the animal through the states and receive a reward.
• Contingency Changes: After 11 rewards, the sequence of correct IAs changed without an explicit cue. Rats had to identify the new $a_1$ and $a_2$ to resume earning rewards. The task structure (states and transitions) remained constant; only the mapping of actions to roles changed.
• Behavioral Analysis: Rats were trained over months to reach a stable performance level. Behavioral metrics included the number of pokes per reward, error types (swap vs. off-sequence), and search strategies (sampling with vs. without replacement).
• Neural Recording: Neuropixels 1.0 probes were chronically implanted in the OFC of three rats. Activity was recorded from untethered, freely moving rats during task performance.
• Data Processing:
  • Motor-related variance (head direction, running speed) was regressed out from neural activity.
  • Pre-poke activity was quantified using a 2-second window preceding each poke.
  • Units were classified as selective for Action (specific IA), State (State 1 vs. State 2), or Role (the function of the action: $a_1$, $a_2$, or off-target).
  • Linear models and decoders were used to quantify unique explained variance and predict task variables from population activity.
  • Population trajectories were analyzed to track learning-dependent changes in neural space relative to time-dependent dynamics.

Key Results

1. Behavioral Few-Shot Learning: Rats demonstrated rapid adaptation, reaching near-optimal performance (approx. 3 pokes per reward) within 2–3 rewards after a contingency change. They utilized the task's recurring structure to generalize across blocks.

   • Learning occurred via state transitions even in the absence of immediate reward; a correct transition from State 1 to State 2 was sufficient to reinforce the action, even if the subsequent State 2 action was incorrect.
   • Rats leveraged knowledge of the task structure: if $a_1$ changed, they were less likely to persist with the previous $a_2$, indicating they inferred a potential change in the entire sequence.
   • Post-learning errors were structured: "swap errors" (performing the correct action for the wrong state) were more frequent than random off-sequence errors, and off-sequence errors clustered near the correct physical location.

2. OFC Encodes Abstract Roles:

   • Approximately 65% of OFC units represented at least one task variable (Action, State, or Role).
   • Crucially, a subset of units encoded Role: the abstract function of an action (e.g., "first correct step" vs. "off-target") independent of the physical identity of the IA. For example, an IA serving as $a_1$ in one block and off-target in the next elicited distinct neural responses corresponding to its new role.
   • These role representations were distinct from value coding; the majority of role-selective units showed non-monotonic response profiles inconsistent with a simple ordinal value hierarchy ($off-target < a_1 < a_2$).

3. Rapid Remapping of Role Representations:

   • The neural representation of role adapted on the same rapid timescale as behavioral learning.
   • Population trajectories preceding critical pokes showed that the neural state associated with a specific action shifted from an "off-target" pattern to the correct "role" pattern within 2–3 critical pokes (rewards).
   • This learning-dependent shift in population activity was nearly orthogonal to the time-dependent dynamics of the trial, indicating a specific reconfiguration of the neural code rather than a slow drift.
   • Decoders trained on role could predict the correct role of an action with high accuracy, and this predictive power emerged rapidly as the rat learned the new sequence.

Significance and Claims
The paper claims that few-shot learning in this task is supported by the rapid remapping of abstract roles onto actions within the OFC.

• Mechanism of Flexibility: The study suggests that the OFC does not merely incrementally update the value of specific actions based on reward history. Instead, it maintains a stable, abstract representation of the task structure (states and roles) and rapidly reassigns specific actions to these roles when contingencies change.
• Temporal Scale: This work demonstrates that abstract representations in the OFC are not static or slow-to-form; they can be updated on the timescale of seconds to support immediate behavioral flexibility.
• Theoretical Implication: The findings support the view that learning involves a transition from detailed, experience-based evaluations to compressed, categorical representations (roles) that capture the invariant structure of a problem, allowing for efficient generalization to new instances.

Limitations
The authors note that the data are correlational; while OFC role representations change in synchrony with behavior, causal necessity was not established (requiring targeted disruption). Additionally, the study analyzed variables separately, though they are inherently coupled, and did not identify specific cell types or projection targets of the recorded units.