Estimating the dimensionality of the manifold underlying multi-electrode neural recordings

Ege Altan, Sara A. Solla, Lee E. Miller, Eric J. PerreaultPLOS Computational Biology 2021 · DOI

TL;DR

Neural recordings are incredibly high-dimensional, but the underlying activity often lies on a lower-dimensional "manifold." This paper rigorously evaluates various methods for estimating this intrinsic dimensionality, revealing their strengths and weaknesses, especially under realistic conditions like noise and non-linear embeddings. They propose a robust pipeline for estimating dimensionality in experimental neural data, crucial for understanding complex brain computations.

What Problem Does This Paper Address?

Simultaneously recording from thousands of neurons in the brain generates massive datasets. However, the activity of these neurons is often highly correlated, meaning the true "degrees of freedom" needed to describe the population's state is much lower than the number of neurons recorded. This underlying structure is often conceptualized as a low-dimensional "manifold" embedded within the high-dimensional space of all possible neural activity patterns. Accurately estimating the dimensionality of this manifold is critical for:

  • Simplifying data analysis: Reducing dimensionality can make it easier to interpret neural activity and identify meaningful patterns.
  • Understanding neural computation: The dimensionality of neural activity might reflect the complexity of the computations performed by a neural circuit.
  • Brain-computer interfaces (BCIs): Lower-dimensional representations can facilitate more efficient and robust control signals for BCIs.

The challenge lies in the fact that real neural data is often noisy and the relationship between the low-dimensional manifold and the high-dimensional recorded signals can be non-linear. Existing dimensionality estimation methods have not been systematically evaluated under these realistic conditions, leading to a lack of consensus on which methods are most suitable for neural data.

Methods

The authors evaluated a range of dimensionality estimation algorithms, broadly categorized into:

  • Linear Algorithms: These assume a linear relationship between the latent low-dimensional space and the high-dimensional observed data. Key methods tested include:
  • Principal Component Analysis (PCA) with a variance cutoff: A common but somewhat arbitrary method.
  • Participation Ratio (PR): A more principled approach to selecting principal components.
  • Parallel Analysis (PA): Compares eigenvalues to a null distribution derived from shuffled data.
  • Nonlinear Algorithms: These can handle more complex, non-linear relationships. Methods included:
  • Correlation Dimension (CD): Estimates dimensionality based on how many points fall within hyperspheres of varying radii.
  • Levina-Bickel Maximum Likelihood Estimation (LBMLE): A maximum likelihood approach to estimate distances.
  • Two Nearest Neighbors (TNN): Uses the ratio of distances to first and second nearest neighbors.
  • Fisher Separability Analysis (FSA): Assesses how easily data points can be separated.

To address the issue of noise, they also developed and tested two denoising approaches:

  • PCA Denoising: A linear method that reconstructs data from principal components.
  • Joint Autoencoder (JAE): A novel deep learning-based nonlinear autoencoder that learns a shared latent representation from two partitions of the data.

Datasets

The study primarily used simulated datasets designed to mimic key properties of real neural recordings. These synthetic datasets allowed for precise control over:

  • Intrinsic dimensionality: The true number of degrees of freedom.
  • Embedding type: Linear or non-linear mapping from latent to observed space.
  • Noise levels: Varying signal-to-noise ratios.
  • Amount of data: Number of samples (time points) and channels.

The simulations were based on empirical firing rate distributions from multi-electrode array recordings in the macaque primary motor cortex (M1) during a center-out task. They also applied their recommended pipeline to three sets of real neural recordings (J1, J2, J3) for validation, as detailed in the Supporting Information.

Key Insights and Ideas

  • Linear vs. Nonlinear Embeddings: Linear algorithms dramatically overestimate dimensionality when the underlying manifold is non-linear, even in the absence of noise. This highlights the critical importance of choosing the right algorithm based on the expected embedding type.
  • Noise is a Major Hurdle: All tested algorithms tend to overestimate dimensionality in the presence of noise, even at relatively high signal-to-noise ratios (e.g., 20 dB). This underscores the necessity of denoising.
  • Denoising is Crucial: Both PCA and the novel Joint Autoencoder significantly improved dimensionality estimates after denoising. The Joint Autoencoder was particularly effective for nonlinear datasets, suggesting its utility for more complex neural activity.
  • Intrinsic Dimensionality Limits: No algorithm performed reliably when the intrinsic dimensionality was high (above ~20) or when the amount of data was very low. This suggests practical limits to what can be reliably estimated.
  • Parallel Analysis as a Good Baseline: For linear datasets or mildly non-linear ones, Parallel Analysis emerged as a robust and computationally efficient linear method, often performing as well as or better than some nonlinear methods.

Key Results

  • On noise-free linear datasets with d=6, all tested linear and nonlinear algorithms accurately estimated the dimensionality.
  • On noise-free nonlinear datasets (α=16, d=6), linear algorithms overestimated dimensionality by over 400%, while nonlinear methods (LBMLE, TNN) performed well.
  • Noise consistently led to overestimation for all algorithms. For example, at SNR=20 dB, nonlinear data dimensionality was overestimated by ~200%.
  • Denoising with PCA improved estimates for linear datasets, while JAE was more effective for nonlinear datasets. After JAE denoising, nonlinear estimates were still overestimated with increasing noise, but at a much slower rate.
  • High intrinsic dimensionality (d > 6) led to underestimation by LBMLE and TNN, with estimates saturating around d=13 for nonlinear data.
  • Amount of data: For linear datasets, PA needed ~20 channels. LBMLE and TNN required ~600 samples (approx. 30s of binned data) for accurate estimates. For nonlinear datasets, LBMLE and TNN needed >50 channels and ~600 samples.

Limitations and Future Work

  • Noise Model Simplification: The study primarily used additive Gaussian isotropic noise. Real neural recordings might have non-additive, non-isotropic, or non-Gaussian noise, which could affect denoising performance.
  • Data Scaling: Simulated data was scaled to [0,1], which might not fully capture the heterogeneity in firing rate ranges observed in real recordings.
  • High Intrinsic Dimensionality: Estimating dimensionality above ~20 proved difficult or impractical for all tested methods.
  • Future Work: The authors suggest investigating more complex noise models and exploring alternative denoising methods (e.g., Latent Factor Analysis through Dynamical Systems - LFADS) that explicitly model temporal dynamics. They also propose that the next step is to investigate the dynamics of signals within the neural manifold and their relation to behavior, building on the dimensionality estimation framework.

Summary updated

BibTeX
@article{altan2021estimating,
  author = {Ege Altan and Sara A. Solla and Lee E. Miller and Eric J. Perreault},
  title = {{Estimating the dimensionality of the manifold underlying multi-electrode neural recordings}},
  journal = {PLOS Computational Biology},
  year = {2021},
  doi = {10.1371/journal.pcbi.1008591},
}
Figure