Continuous multiplexed population representations of task context in the mouse primary visual cortex

Márton A. Hajnal, Duy Tran, Michael Einstein, Mauricio Vallejo Martelo, Karen Safaryan, Pierre‐Olivier Polack, Peyman Golshani, Gergő OrbánbioRxiv 2021 · DOI

TL;DR

This study reveals that the primary visual cortex (V1) in mice doesn't just process raw visual information; it also actively represents abstract cognitive variables like "task context." Remarkably, this context information is encoded in a way that is largely independent of the visual stimuli themselves, suggesting V1 plays a more sophisticated role in decision-making than previously thought.

What Problem Does This Paper Address?

The brain needs to interpret sensory information flexibly, adapting its response based on the current situation or "task context." For instance, the same visual cue might mean "look closer" in one scenario and "ignore" in another. While higher brain areas are known to handle such contextual processing, it's been unclear whether primary sensory areas like V1, the first stop for visual information, also contribute to this. Specifically, the authors investigate whether V1 neurons can represent and encode task context, and if so, how this representation relates to the processing of the visual stimuli themselves. They aim to understand if V1 can disentangle and represent abstract rules of engagement from the raw sensory input.

Methods

The researchers trained mice to perform a challenging cross-modal decision task where the interpretation of identical audio-visual stimuli depended on the task context. They used multi-channel silicon probes to record neuronal activity from V1 in awake, behaving mice. Key analytical techniques included:

  • Population Decoding: Using linear decoders to determine if task context (e.g., whether to attend to visual or auditory cues) and visual stimulus identity could be predicted from the collective activity of V1 neurons.
  • Principal Component Analysis (PCA): To identify the principal dimensions of variance in population activity and understand how different task variables contribute to this variance.
  • Multidecoder Subspace Analysis (MDSA): A novel approach to analyze the relationship between representations of different task variables by projecting population activity onto a subspace defined by multiple decision vectors. This allowed them to assess the orthogonality (independence) of context and visual stimulus representations.
  • Time-Shifted Decoding: To examine the temporal dynamics of context representation and how it changes across different phases of the trial (pre-stimulus, stimulus presentation, post-stimulus).
  • Locomotion Matching Analysis: To control for potential confounds from movement-related activity by matching running speed distributions across different task conditions.

Datasets

The study used data collected from 15 mice performing a custom-designed cross-modal decision-making task. No pre-existing datasets were used for training or evaluation.

Key Insights and Ideas

  • V1 Encodes Abstract Context: The most striking finding is that V1 neurons can represent abstract task context, not just raw sensory input. This means V1 is not merely a passive relay station but actively participates in higher-level cognitive processing.
  • Orthogonal Representations: Task context and visual stimulus identity are represented in largely orthogonal subspaces within V1 population activity. This is crucial because it suggests that V1 can represent these two types of information independently, preventing interference and allowing for flexible interpretation of stimuli.
  • Dynamic Context Representation: Context representation isn't static; it shows distinct dynamic shifts, particularly at stimulus onset and offset, indicating an active and adaptive neural process.
  • Context in the Absence of Stimuli: Task context can be decoded not only during stimulus presentation but also during the intertrial intervals (off-stimulus periods), suggesting V1 maintains this information even without direct sensory input.
  • Mixed Selectivity: While representations are orthogonal, the underlying neuronal populations show "mixed selectivity," meaning individual neurons can be sensitive to multiple task variables, contributing to a rich and flexible coding scheme.

Key Results

  • Task context could be decoded with high accuracy (around 65-68%) from V1 population activity during both stimulus presentation and intertrial intervals.
  • The representation of task context was found to be orthogonal to the representation of visual stimulus identity in population activity space.
  • Context representation was dynamic, showing significant shifts upon stimulus onset and offset.
  • Locomotion matching analysis confirmed that the identified task context representation was not confounded by variations in locomotion patterns.

Limitations and Future Work

  • Input Pathways: The paper acknowledges that the precise neural pathways through which contextual information reaches V1 are not fully elucidated. Future work could involve recordings from projecting neurons and manipulations of their activity.
  • Mechanism of Transformation: While the dynamic transformation of context representation at stimulus onset/offset is observed, the precise mechanisms driving this rapid change are not fully detailed. Further investigation into population-wide transformations and neuronal recruitment is suggested.
  • Role in Perception/Decision: The study demonstrates context representation in V1 but doesn't fully detail how this representation directly contributes to perception and decision-making, especially at the edge of perceptual thresholds. Future work could explore this link.
  • Movement-Related Activity: While locomotion was controlled for, the authors note that other movement-related activities not captured by running speed might still contribute to confounds. Further analysis to disentangle these is suggested.

Summary updated

Abstract

Summary Primary visual cortex (V1) neurons integrate motor and multisensory information with visual inputs during sensory processing. However, whether V1 neurons also integrate and encode higher-order cognitive variables is less understood. We trained mice to perform a context-dependent cross-modal decision task where the interpretation of identical audio-visual stimuli depends on task context. We performed silicon probe population recordings of neuronal activity in V1 during task performance and showed that task context (whether the animal should base its decision on visual or auditory stimuli) can be decoded during both intertrial intervals and stimulus presentations. Context and visual stimuli were represented in overlapping populations but were orthogonal in the population activity space. Context representation was not static but displayed distinctive dynamics upon stimulus onset and offset. Thus, activity patterns in V1 independently represent visual stimuli and cognitive variables relevant to task execution.
BibTeX
@misc{hajnal2021continuous,
  author = {Márton A. Hajnal and Duy Tran and Michael Einstein and Mauricio Vallejo Martelo and Karen Safaryan and Pierre‐Olivier Polack and Peyman Golshani and Gergő Orbán},
  title = {{Continuous multiplexed population representations of task context in the mouse primary visual cortex}},
  year = {2021},
  doi = {10.1101/2021.04.20.440666},
  howpublished = {bioRxiv},
}
Figure