Misinterpreting the horseshoe effect in neuroscience

Timothée Proix, Matthew G. Perich, Tomislav MilekovicbioRxiv 2022 · DOI

TL;DR

Dimensionality reduction techniques often reveal "horseshoe" shaped, oscillatory trajectories in neural data, which have been interpreted as evidence of underlying latent dynamical systems. This paper demonstrates that these trajectories are an artifact of the analysis method itself, arising whenever data exhibits continuous variation over time, regardless of the true underlying dynamics. This finding suggests that these common oscillatory patterns in neuroscience should not be automatically interpreted as evidence of latent dynamics.

What Problem Does This Paper Address?

Neuroscientists frequently analyze high-dimensional neural recordings (e.g., from hundreds or thousands of neurons) to understand brain activity during behavior. A common challenge is to visualize and interpret this complex data. Dimensionality reduction techniques are widely used to project this high-dimensional activity into a lower-dimensional space, often revealing trajectories that evolve over time. A recurring observation in these analyses is the emergence of "horseshoe" shaped, oscillatory trajectories. These patterns have been interpreted as strong evidence for the existence of underlying "latent dynamical systems" – internal states or processes that govern neural activity. However, the authors question whether these observed trajectories are truly indicative of these latent dynamics or if they might be an artifact of the analysis methods themselves.

Methods

The core of the paper's argument relies on the "horseshoe effect," a mathematical principle. The authors analytically demonstrate that applying common dimensionality reduction methods (like PCA, multidimensional scaling, etc.) to data exhibiting continuous variation in time will necessarily produce oscillatory trajectories. They illustrate this with a novel analogy: observing neuroscientists at a conference. As an observer walks through a hallway divided by nationality, the number of neuroscientists from each country changes continuously. Applying PCA to this data, even though there are no inherent "dynamics" in the neuroscientists' grouping, results in oscillatory latent variables.

They further support this by simulating neural spiking activity. By creating a model where neuron firing rates are tuned to a time instant (mimicking continuous variation), they show that dimensionality reduction on this simulated data also yields oscillatory trajectories, regardless of whether the underlying model incorporates actual dynamical systems or simply delayed connections. Finally, they analyze real neural recordings from monkeys performing motor and visual tasks, demonstrating that these datasets exhibit continuous variation and, when analyzed with dimensionality reduction, produce oscillatory trajectories consistent with the horseshoe effect.

Datasets

The paper utilizes several real neural datasets:

  • Motor cortex (M1) and premotor dorsal cortex (PMd) data: From a rhesus macaque monkey performing a center-out reaching task.
  • Primary somatosensory cortex (S1) data: From another rhesus macaque monkey performing a similar reaching task.
  • V4 visual cortex data: From a third rhesus macaque monkey performing a fixation task with visual stimuli.

These datasets are from previously published studies and are available upon request or from public repositories (Dryad, CRCNS.org). The paper also introduces simulated datasets generated using their models to demonstrate the horseshoe effect.

Key Insights and Ideas

  • The "horseshoe effect" is a powerful explanation for oscillatory neural trajectories: The paper provides a strong mathematical and empirical argument that these common patterns are not necessarily evidence of latent dynamics but rather an artifact of dimensionality reduction applied to time-varying data.
  • Continuous variation is the key driver: The presence of continuous variation in the data, not necessarily complex internal dynamics, is sufficient to generate oscillatory trajectories when dimensionality reduction is applied.
  • Reinterpreting existing findings: This work suggests that many previous interpretations of neural data, which relied on oscillatory trajectories as evidence for latent dynamical systems, may need to be re-evaluated.
  • Simulations are crucial for understanding analysis artifacts: The use of both a simple analogy and simulated neural data effectively isolates the effect of the dimensionality reduction method from the underlying biological system.

Key Results

  • Analytical proof: The paper analytically shows that dimensionality reduction on data with continuous variation will produce oscillatory latent variables.
  • Simulation validation: Simulated neural data exhibiting continuous variation, when subjected to dimensionality reduction, yielded oscillatory trajectories that closely resembled those observed in real neural data.
  • Real data analysis: Analysis of multiple real neural datasets from monkeys performing motor and visual tasks confirmed that these datasets exhibit continuous variation and that dimensionality reduction produces oscillatory trajectories consistent with the horseshoe effect. For instance, in the M1 dataset, the temporal covariance matrix was well approximated by a Toeplitz matrix (indicating continuous variation), and the resulting latent variables traced oscillatory trajectories.
  • Robustness to noise and sampling: Simulations showed that the horseshoe effect persists even with noise, sparse sampling, or biased sampling in the data.

Limitations and Future Work

  • Limitations: The paper acknowledges that real-life neural datasets will rarely exhibit ideal continuous variation due to noise, random sampling, or bias. This means the temporal covariance matrix will deviate from the ideal Toeplitz structure. However, they show that the similarity of the trajectories to the horseshoe curve remains proportional to the similarity of the covariance matrix to the Toeplitz structure, suggesting the effect is still dominant.
  • Future Work: The authors emphasize the critical need to complement dimensionality reduction analysis with other methods to avoid misinterpreting artifacts as biological dynamics. They suggest analyzing inter-conditional relationships to reveal latent organization independently of dynamics and verifying any uncovered features through causal experimental manipulations. Ultimately, they propose that identifying latent dynamics in high-dimensional neural data might be infeasible if the data inherently exhibits continuous variation.

Summary updated

Abstract

ABSTRACT Dimensionality reduction methods are frequently used to analyze high-dimensional activity of cortical neuron populations during behavior. The resulting oscillatory trajectories that consistently emerge from this analysis have been interpreted as a signature of latent dynamical systems. Here, we show that these oscillatory trajectories necessarily result from applying dimensionality reduction methods on recordings that approximately exhibit continuous variation in time, regardless of whether or not the recorded system incorporates latent dynamics.
BibTeX
@misc{proix2022misinterpreting,
  author = {Timothée Proix and Matthew G. Perich and Tomislav Milekovic},
  title = {{Misinterpreting the horseshoe effect in neuroscience}},
  year = {2022},
  doi = {10.1101/2022.03.04.482986},
  howpublished = {bioRxiv},
}

Loading related content...

Figure