Recycling of prefrontal subspaces dynamically multiplexes information

Pablo T Wentz, Scott L. Brincat, Anitha Pasupathy, Earl K. MillerbioRxiv 2025 · DOI

TL;DR

The prefrontal cortex (PFC) doesn't just use distinct neural "subspaces" for different tasks or time points; it dynamically reuses and repurposes these subspaces. This study shows that the same neural subspace can be reactivated for the same information later in a trial or even recycled to represent entirely new types of information, suggesting PFC operates more like a flexible "workspace" than a fixed information processor.

What Problem Does This Paper Address?

Neuroscientists have long understood that the brain encodes information not in individual neurons, but in the collective activity patterns of neural populations. Within these populations, information is often organized into lower-dimensional "subspaces." A common assumption has been that these subspaces are specialized: a given subspace is dedicated to a specific type of information (e.g., sensory input) and is active only during a particular time window (e.g., encoding). This paper challenges this assumption by investigating the flexibility of these PFC subspaces, asking if they are truly fixed or if they can be dynamically repurposed.

Methods

The researchers employed a sophisticated analysis technique called demixed principal component analysis (dPCA). This method is designed to identify low-dimensional subspaces within neural population activity that are maximally informative about specific task variables (like cue identity, location, or response direction) while minimizing variance related to other variables.

They then used a cross-temporal generalization approach. This involves training the dPCA model on data from one time point and then testing how well that subspace captures information at different time points. This allows them to track how information is represented and whether subspaces are reactivated or reused across time.

Finally, they used cross-factor generalization to see if a subspace trained to represent one type of information (e.g., cue location) could also encode a different type of information (e.g., response direction) at a later time.

Datasets

The study utilized two datasets from non-human primates (NHPs) performing visuomotor association tasks:

  1. Object Task: This task involved integrating object identity with a learned rule to generate a behavioral response. Data was collected from 632 neurons in two NHPs.
  2. Object/Location Task: This task required integrating object identity and spatial location to generate a behavioral response. Data was collected from 509 neurons in one NHP.

Key Insights and Ideas

  • Subspace Reactivation: The same neural subspace used for initial information encoding can be reactivated later in the trial to represent the same information, specifically for generating a behavioral response. This challenges the idea that subspaces are only active during their initial encoding phase.
  • Subspace Recycling: Subspaces initially dedicated to processing sensory information (like cue location or object identity) can be "recycled" to represent entirely different types of information, such as the behavioral response direction.
  • Flexible "Workspaces": These findings suggest that PFC subspaces function more like flexible "workspaces" where information can be written, read, and rewritten, rather than being rigidly assigned to specific information or time windows.
  • Preservation of Spatial Mapping: When subspaces are recycled, their original spatial organization is often preserved. For example, a subspace that encoded "left vs. right" cue locations can later encode "left vs. right" motor responses in a consistent manner.
  • Non-spatial Subspaces Can Be Recycled: Even subspaces that initially encode non-spatial information (like cue object identity) can be recycled to represent spatial information (like location and response direction), demonstrating a broad flexibility in PFC coding.

Key Results

  • In the Object Task, the encoding subspace for response direction was reactivated during the behavioral response epoch, carrying the same information.
  • In the Object/Location Task, subspaces initially representing cue location were later recycled to encode response direction, and cue object subspaces were recycled to encode both cue location and response direction.
  • The study found that these recycled subspaces often maintained their original spatial mapping, meaning "left" in the original context corresponded to "left" in the recycled context.
  • The results indicate that PFC subspaces are not rigidly tied to specific information types or time windows but are dynamically deployed and repurposed.

Limitations and Future Work

  • NHP Model: The study was conducted on non-human primates. While valuable, generalizing these findings directly to human cognition requires further investigation.
  • Task Specificity: The observed subspace recycling and reactivation were demonstrated in specific, albeit complex, visuomotor association tasks. It remains to be seen how broadly these principles apply to other cognitive functions.
  • Distinguishing Hypotheses: The authors propose two hypotheses for why subspace recycling might be more prominent in the Object/Location task compared to the Object task: (1) the strength of learned associations, or (2) the requirement to bind object and location information. Distinguishing between these will require further experimentation.
  • Future Directions: The authors suggest that future models of PFC function, particularly recurrent neural networks, should be updated to account for subspace reactivation and recycling. They also propose that understanding the computations enabled by these flexible subspaces, rather than just the information they represent, is crucial. Further research could explore how these flexible workspaces are utilized in other cognitive domains and how their structure is maintained or altered.

Summary updated

Abstract

Abstract Cortical spiking activity in areas like prefrontal cortex (PFC) encodes information only along a small number of dimensions, a subspace of the full space of population activity patterns. PFC often uses distinct subspaces to represent incoming sensory information and to maintain it in working memory. However, it’s generally assumed that a subspace carries specific information within a specific time frame. Here, in two datasets, we show that neither assumption is correct. The same subspace used during initial encoding was later reactivated during the behavioral response. We also found that subspaces ostensibly dedicated to processing sensory information were recycled to represent behavioral response information later in a trial. We further found a consistent spatial mapping between the original subspace usage and its later recycling. These results demonstrate previously unknown flexibility of PFC subspace coding, both in the deployment of subspaces across time and in their multiplexing of information. This flexibility suggests that PFC subspaces function more like “workspaces” into which different types of information can be written.
BibTeX
@misc{wentz2025recycling,
  author = {Pablo T Wentz and Scott L. Brincat and Anitha Pasupathy and Earl K. Miller},
  title = {{Recycling of prefrontal subspaces dynamically multiplexes information}},
  year = {2025},
  doi = {10.1101/2025.05.02.651965},
  howpublished = {bioRxiv},
}
Figure