Population coding of strategic variables during foraging in freely moving macaques

Neda Shahidi, Melissa Franch, Arun Parajuli, Paul Schrater, Anthony A. Wright, Xaq Pitkow, Valentin DragoiNature Neuroscience 2024 · DOI

TL;DR

Researchers have finally peered into the "free-range" brain of foraging primates. By wirelessly recording neural activity in unrestrained macaques, the study reveals that monkeys don't just react to rewards; they maintain a sophisticated internal model of reward dynamics—tracking both reward history and the passage of time—encoded within a specific, low-dimensional subspace of the prefrontal cortex.

What Problem Does This Paper Address?

Most of what we know about the neural basis of decision-making comes from "trial-based" experiments where animals are physically restrained and forced to choose between options at specific intervals. This is a far cry from nature, where foraging is a continuous process of deciding both where to go and when to try.

Physical restraint and rigid trial structures distort an animal's perception of time and spatial reward distribution. The authors argue that to truly understand the neural basis of foraging, we must study animals that are free to move and choose their own pace. The core question: How does the brain represent the "strategic variables" (like waiting time and reward history) that guide behavior in a naturalistic, continuous environment?

Methods

The team used a "trial-free" foraging task where two macaques could move freely between two reward boxes. Rewards were delivered on a Variable-Interval (VI) schedule, meaning the longer a monkey waited to press a button, the higher the probability a reward would be available (though this was hidden from the monkey).

Key technical highlights include:

  • Wireless Electrophysiology: They used a lightweight, wireless transmitter to record from a 96-channel Utah array implanted in the dorsolateral prefrontal cortex (dlPFC).
  • Locomotion Tracking: An overhead camera and DeepLabCut were used to track movement, allowing researchers to "decorrelate" neural activity related to walking from activity related to decision-making.
  • Canonical Correlation Analysis (CCA): To find the signal in the noise, they used CCA to identify a "latent space"—a compressed, low-dimensional version of the high-dimensional neural population activity that specifically represented task-relevant variables.
  • Decoding: They used logistic regressions and Generalized Linear Models (GLMs) to see if neural activity could predict the monkey's next move.

Datasets

The study collected data from:

  • Two primary macaques (G and T): Used for the main foraging task and neural recordings (30 sessions total, 1,323 single and multi-units).
  • One control macaque (M): Used specifically to monitor eye movements, pupil diameter, and limb movements to ensure these "task-irrelevant" variables weren't confounding the dlPFC data.
  • Behavioral Data: A pool of 8,862 individual button presses across various reward schedules.

Key Insights and Ideas

  • The "When" Matters as Much as the "Where": Unlike traditional "matching law" theories which suggest animals just follow the highest reward rate, these monkeys showed they understood the dynamics of the task. They knew that waiting longer improved their odds, adjusting their "response rate" based on their subjective expectation of reward.
  • Subjective Beliefs > Objective Reality: The most striking finding was that the neural representation of task variables (the "latent components") predicted the monkey's future choices better than the actual, objective task variables themselves. This suggests the dlPFC is encoding the animal's internal model or "guess" of the world, which is what actually drives behavior.
  • Neural Ramping: The neural component representing "waiting time" showed a ramping activity between presses. This looks remarkably like "evidence accumulation" models often seen in simpler laboratory tasks, but here it was happening spontaneously in a free-moving animal.
  • Strategic Compression: The brain doesn't use all its neurons independently for these tasks. Instead, it compresses the complex information into a specific subspace. This "targeted dimensionality reduction" revealed the underlying computations more clearly than looking at raw neural firing or behavior alone.

Key Results

  • High Correlation: The probability of reward availability was highly correlated with the monkeys' behavior ($r = 0.93$).
  • Choice Prediction: Neural activity in the dlPFC significantly predicted whether a monkey would "stay" at a box or "switch" to the other one before the action occurred (AUC ~0.7-0.8).
  • Waiting Time Prediction: The neural population could predict exactly how long the monkey would wait before its next press ($r \approx 0.4-0.5$ in some sessions).
  • Subspace Efficiency: The 10-dimensional neural subspace identified via CCA captured the relevant signals for subsequent choices as effectively as the entire recorded population, proving that foraging strategies are low-dimensional.

Limitations and Future Work

The authors acknowledge that their reward dynamics were stochastic and time-based (resembling how fruit might ripen or prey might return to a patch). It remains to be seen if these findings hold for other reward schemes, such as "count-based" rewards (where a reward appears after $X$ number of attempts regardless of time).

Additionally, while the dlPFC is a major hub for these variables, foraging is a whole-brain effort. Future work could expand wireless recordings to other areas like the anterior cingulate cortex (ACC) or the basal ganglia to see how this internal model is distributed and transformed into motor actions.

Summary updated

Abstract

Until now, it has been difficult to examine the neural bases of foraging in naturalistic environments because previous approaches have relied on restrained animals performing trial-based foraging tasks. Here we allowed unrestrained monkeys to freely interact with concurrent reward options while we wirelessly recorded population activity in the dorsolateral prefrontal cortex. The animals decided when and where to forage based on whether their prediction of reward was fulfilled or violated. This prediction was not solely based on a history of reward delivery, but also on the understanding that waiting longer improves the chance of reward. The task variables were continuously represented in a subspace of the high-dimensional population activity, and this compressed representation predicted the animal's subsequent choices better than the true task variables and as well as the raw neural activity. Our results indicate that monkeys' foraging strategies are based on a cortical model of reward dynamics as animals freely explore their environment.
BibTeX
@inproceedings{shahidi2024population,
  author = {Neda Shahidi and Melissa Franch and Arun Parajuli and Paul Schrater and Anthony A. Wright and Xaq Pitkow and Valentin Dragoi},
  title = {{Population coding of strategic variables during foraging in freely moving macaques}},
  booktitle = {Nature Neuroscience},
  year = {2024},
  doi = {10.1038/s41593-024-01575-w},
}
Figure