Are task representations gated in macaque prefrontal cortex?

Timo Flesch, Valerio Mante, William T. Newsome, Andrew Saxe, Christopher Summerfield, David SussilloarXiv 2023 · DOI · arXiv

Abstract

A recent paper (Flesch et al, 2022) describes behavioural and neural data suggesting that task representations are gated in the prefrontal cortex in both humans and macaques. This short note proposes an alternative explanation for the reported results from the macaque data.

Introduction

A central question in systems neuroscience is how the brain selectively processes relevant information while ignoring distractors. In context-dependent tasks—such as choosing a stimulus based on its color rather than its shape—the mechanism for "gating" irrelevant features is debated.

A 2022 study by Flesch et al. (FSS) proposed that both humans and macaques employ a "compression" strategy, where neural dimensions representing irrelevant features are squashed in the prefrontal cortex. This finding, however, challenged the landmark 2013 study by Mante, Sussillo, and Newsome (MSN), which found that in the macaque Frontal Eye Fields (FEF), irrelevant information remains fully represented but is ignored during the final decision stage. This note re-examines the macaque data to determine if these representations are truly gated or if the observed compression is a statistical artifact.

Methods

The authors investigated the discrepancy by comparing the analysis pipelines and underlying models of the FSS and MSN studies. Two primary technical differences were identified: the frame of reference for data analysis and the handling of motor-related signals (orthogonalization).

Comparison of Mante coding (early transform) and Flesch coding (late transform) architectures for context-dependent tasks.

As illustrated in Figure 3, the models differ in where the transformation from sensory input to motor response occurs. "Mante coding" (MSN) utilizes an early transform where sensory inputs are aligned to the response frame before entering the recurrent neural network (ANN). In contrast, "Flesch coding" (FSS) processes raw sensory inputs in the ANN and applies a late transform to map choices to physical responses. The authors used simulations to test whether an uncompressed MSN-style model would appear compressed when analyzed using the FSS pipeline.

Key Results

The re-analysis revealed that the evidence for compression in macaques reported by FSS was likely spurious, arising from two main factors:

  1. Frame of Reference Artifacts: Because motor signals in the FEF are dominant, failing to account for the response-aligned nature of the task can make sensory signals appear compressed when they are actually just overshadowed by the motor plan.
  2. Simulation Match: Simulations of an uncompressed MSN architecture, when processed through the FSS analysis pipeline, successfully reproduced the appearance of compression. This confirms that the FSS method can detect compression even in its absence.
  3. Restored Consistency: When these analytical differences are resolved, the macaque FEF data remains consistent with the original 2013 conclusion: input representations are largely uncompressed and are filtered at a late stage of selection.

Discussion

While this note clarifies the macaque data, it highlights a persistent "species gap." Human fMRI data continues to show evidence of early gating and compression, whereas macaque single-unit recordings do not.

Several factors may explain this divergence. First, the recording methods differ; fMRI measures aggregate blood flow which may inherently smooth or "compress" signals compared to individual neuronal spikes. Second, the brain regions studied were different: the macaque data focused on the motor-centric FEF, while the human data covered broader prefrontal and parietal areas. Finally, it remains possible that humans and macaques utilize fundamentally different neural geometries to solve similar cognitive problems. Future research across species using comparable recording techniques and brain regions will be necessary to resolve this gap.

Summary updated

Abstract

A recent paper (Flesch et al, 2022) describes behavioural and neural data suggesting that task representations are gated in the prefrontal cortex in both humans and macaques. This short note proposes an alternative explanation for the reported results from the macaque data.
BibTeX
@misc{flesch2023are,
  author = {Timo Flesch and Valerio Mante and William T. Newsome and Andrew Saxe and Christopher Summerfield and David Sussillo},
  title = {{Are task representations gated in macaque prefrontal cortex?}},
  year = {2023},
  eprint = {2306.16733},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.16733},
}
Figure