Measuring Fisher Information Accurately in Correlated Neural Populations

Ingmar Kanitscheider, Ruben Coen Cagli, Adam Kohn, Alexandre PougetPLOS Computational Biology 2015 · DOI

TL;DR

This paper introduces a novel, computationally efficient, and statistically unbiased method for measuring Fisher information in neural populations, even when neural responses are correlated. This new approach significantly outperforms traditional decoding-based methods, especially with limited experimental data, offering a more accurate and faster way to quantify neural coding reliability.

What Problem Does This Paper Address?

Understanding how the brain represents sensory information relies on quantifying the reliability of neural populations. Neural responses are inherently noisy and variable, and crucially, the noise across neurons is often correlated. These "noise correlations" can significantly impact how much information a population can convey about a stimulus. Traditional methods for measuring this reliability, like cross-validated decoding, are computationally intensive and tend to underestimate the true information, especially when the number of experimental trials is limited – a common scenario in neuroscience. This underestimation is a major bottleneck for accurately comparing coding across different conditions or monitoring changes over time. The paper addresses the need for a more accurate, efficient, and less data-hungry method to measure Fisher information in the presence of these challenging noise correlations.

Methods

The core contribution is a bias-corrected direct estimator for Fisher information. Instead of training a decoder to estimate the stimulus from neural activity, this method directly estimates the Fisher information using the measured tuning curves (how neuron responses change with stimulus) and the noise covariance matrix (how noise in different neurons is related). The key innovation lies in analytically deriving and correcting for the biases that arise when estimating these quantities from finite data. This involves correcting for the non-linear transformations (like inversion and squaring) inherent in the Fisher information formula. The authors also derive an analytical expression for the variance of their estimator, allowing for accurate error bar calculation. For specific cases like measuring information in uncorrelated populations or when ignoring correlations, they propose further extensions of this bias-correction framework.

Datasets

The authors evaluate their method on two types of datasets:

  • Simulated data: Generated from a model of orientation-selective neurons (similar to those in primate V1) with realistic noise correlations. This allowed them to compute the "ground truth" Fisher information and compare their estimator's accuracy against it.
  • Cortical data: Recordings from 52 macaque V1 neurons responding to gratings. This real-world data helps validate the robustness of their method to non-Gaussian response variability.

Key Insights and Ideas

  • Direct estimation is superior to decoding for limited data: When the number of trials is not vastly larger than the number of neurons, directly estimating Fisher information with bias correction is significantly more accurate and less biased than cross-validated decoding.
  • Noise correlations are a major challenge for decoding: Traditional decoding methods struggle to accurately account for noise correlations, leading to underestimation of information, especially when correlations are not explicitly modeled.
  • Analytical bias correction is powerful: By understanding the statistical properties of estimated tuning curves and covariance matrices, it's possible to derive precise corrections for estimation biases, leading to a more accurate Fisher information estimate.
  • Computational efficiency: The direct estimation method is orders of magnitude faster than decoding methods, requiring only matrix inversion rather than iterative optimization.
  • Robustness to non-Gaussianity: The bias-corrected estimator performs well even on real neural data, which often exhibits non-Gaussian response statistics, suggesting its broad applicability.

Key Results

  • The bias-corrected direct estimator closely approximates the true Fisher information even with a limited number of trials (e.g., 250 trials for 50 neurons).
  • For 250 trials, the bias-corrected estimate was within 11% of the true information, while the decoder estimate was within 24%. To achieve similar accuracy, the decoder required 1000 trials.
  • The direct estimator is generally unbiased, whereas cross-validated decoding systematically underestimates information.
  • The computational speedup of the direct estimator over decoding methods is substantial, often by orders of magnitude.
  • The method successfully measures context-dependent encoding and the impact of correlations on information encoding and decoding, outperforming decoding-based approaches in these scenarios as well.

Limitations and Future Work

  • Minimum trial requirement: The direct estimator requires a minimum number of trials, T > (N+5)/2, where N is the number of neurons. If fewer trials are available, decoding remains the only option, though estimates may be inaccurate.
  • Fine discrimination bias: For very small stimulus differences (fine discrimination), the estimation of the tuning curve derivative can be noisy, impacting accuracy. Conversely, for very large differences (coarse discrimination), non-linear tuning curves can introduce bias. The choice of stimulus difference requires a tradeoff.
  • Future work: The authors suggest extending their approach to multiclass classification tasks and developing methods for estimating full (non-linear) Fisher information, which is currently more challenging. They also highlight the need for methods that can handle cases with even fewer trials than currently required by their direct estimator.

Summary updated

BibTeX
@article{kanitscheider2015measuring,
  author = {Ingmar Kanitscheider and Ruben Coen Cagli and Adam Kohn and Alexandre Pouget},
  title = {{Measuring Fisher Information Accurately in Correlated Neural Populations}},
  journal = {PLOS Computational Biology},
  year = {2015},
  doi = {10.1371/journal.pcbi.1004218},
}
Figure