Contribution of behavioural variability to representational drift

Sadra Sadeh, Claudia Clopath

TL;DR

This paper reveals that changes in an animal's behavior, like arousal or movement, significantly contribute to what was previously thought to be "representational drift" in the brain. By analyzing neural activity alongside detailed behavioral data, the researchers show that this behavioral variability can be misinterpreted as drift if not properly accounted for, suggesting a more complex, multidimensional view of neural representations.

What Problem Does This Paper Address?

The brain's ability to represent the external world relies on stable neural activity patterns. However, studies have observed "representational drift," where these patterns change over time even when the stimulus remains the same. While this drift has often been attributed to internal neural plasticity or stimulus-related factors, it's also known that an animal's behavioral state (e.g., arousal, attention, movement) can strongly influence neural activity. The key question this paper tackles is: how does the variability in these behavioral states contribute to representational changes, and can it be mistaken for representational drift itself?

Methods

The researchers employed a multi-pronged approach:

  • Large-scale Electrophysiology Data Analysis: They analyzed publicly available datasets from the Allen Brain Observatory, which include simultaneous recordings of neural activity (using Neuropixels probes) and behavioral parameters (pupil size, running speed) from mice.
  • Representational Similarity Analysis (RSA): They quantified how similar neural responses were across repeated presentations of the same stimulus. This was done by calculating correlation coefficients between population activity vectors.
  • Computational Modeling: They developed and compared two models:
  • Signal-Gain Model: Assumes behavior scales the stimulus-driven signal.
  • Independent-Mixing Model: Assumes behavior and stimulus contribute independently to neural activity.
  • Decoding Analysis: They trained linear decoders to discriminate between different stimulus frames, first using all recorded units and then focusing on "reliable" units (those with consistent stimulus responses). They also trained decoders to predict behavioral states from neural activity.
  • Principal Component Analysis (PCA): Used to identify low-dimensional patterns in population activity and their relationship to behavior.

Datasets

The study primarily utilized publicly available datasets from the Allen Brain Observatory:

  • Neuropixels Dataset 1 (Brain Observatory): 32 sessions/mice.
  • Neuropixels Dataset 2 (Functional Connectivity): 26 sessions/mice.

These datasets are characterized by standardized visual stimulation, large-scale neural recordings, and systematic measurements of behavioral parameters.

Key Insights and Ideas

  • Behavioral Variability is a Major Driver of "Drift": The most striking finding is that changes in an animal's behavioral state (pupil size, running speed) significantly contribute to changes in neural representations, often to an extent that can be misinterpreted as representational drift.
  • Independent Modulation: The data strongly supports an "independent-mixing" model, where behavior modulates neural activity independently of the stimulus, rather than simply scaling the stimulus response (gain model).
  • Low-Dimensional Behavioral Signals: Behavioral state changes are reflected in the low-dimensional components of population activity, suggesting a global influence on neural processing.
  • Misinterpretation of Drift: If neural representations are analyzed solely in the stimulus space, ignoring behavioral variability, then the independent modulation by behavior can be mistaken for representational drift.
  • Reliable Units vs. Behaviorally Modulated Units: While reliable stimulus-encoding units are crucial for stable decoding, focusing solely on them might not be sufficient. Behaviorally modulated units also play a significant role, and their contribution is distinct from stimulus reliability.

Key Results

  • Representational similarity between repeated stimulus presentations was significantly lower when behavioral states (pupil size, running speed) changed more drastically between presentations.
  • Computational models showed that an independent mixing of stimulus and behavioral signals better explained the observed data than a simple gain model.
  • Individual neurons' stimulus reliability and their modulation by behavior were largely independent, supporting the independent-mixing model.
  • Decoding accuracy for natural images decreased significantly when using all units but improved when focusing on units with high stimulus reliability, suggesting behavioral "noise" hinders generalization.
  • Behavioral states (pupil width, running speed) could be reliably predicted from neural activity.

Limitations and Future Work

  • Timescales: The current study focused on changes within a day. The authors acknowledge that longer timescales (days to weeks) might involve different mechanisms, and further studies with more comparison time points are needed.
  • Multidimensionality: The authors emphasize that representational similarity should be understood in a multidimensional parameter space that includes both external (stimulus) and internal (behavioral) parameters.
  • Other Behavioral Parameters: While pupil size and running speed were analyzed, other behavioral parameters (e.g., whisking, posture) were not quantified but could also contribute to representational changes.
  • Causal Interrogation: Future work could leverage behavioral perturbations and large-scale neural recordings to causally investigate the mechanisms underlying representational drift and the role of behavior.
  • Circuit Mechanisms: Understanding the specific circuit mechanisms that mediate the bidirectional modulation of neural activity by behavior is an important future direction.

Summary updated

Abstract

Neuronal responses to similar stimuli change dynamically over time, raising the question of how internal representations can provide a stable substrate for neural coding. Recent work has suggested a large degree of drift in neural representations even in sensory cortices, which are believed to store stable representations of the external world. While the drift of these representations is mostly characterized in relation to external stimuli, the behavioural state of the animal (for instance, the level of arousal) is also known to strongly modulate the neural activity. We therefore asked how the variability of such modulatory mechanisms can contribute to representational changes. We analysed large-scale recording of neural activity from the Allen Brain Observatory, which was used before to document representational drift in the mouse visual cortex. We found that, within these datasets, behavioural variability significantly contributes to representational changes. This effect was broadcasted across various cortical areas in the mouse, including the primary visual cortex, higher order visual areas, and even regions not primarily linked to vision like hippocampus. Our computational modelling suggests that these results are consistent with independent modulation of neural activity by behaviour over slower timescales. Importantly, our analysis suggests that reliable but variable modulation of neural representations by behaviour can be misinterpreted as representational drift if neuronal representations are only characterized in the stimulus space and marginalized over behavioural parameters.
BibTeX
@inproceedings{sadeh2022contribution,
  author = {Sadra Sadeh and Claudia Clopath},
  title = {{Contribution of behavioural variability to representational drift}},
  booktitle = {eLife},
  year = {2022},
  doi = {10.7554/elife.77907},
}
Figure