Extracting information in spike time patterns with wavelets and information theory

Vítor Lopes‐dos‐Santos, Stefano Panzeri, Christoph Kayser, Mathew E. Diamond, Rodrigo Quian QuirogaJournal of Neurophysiology 2014 · DOI

TL;DR

This paper introduces the Wavelet-Information (WI) method, a novel approach that leverages wavelet decomposition and information theory to extract meaningful information encoded in the precise timing of neural spikes. The WI method offers a more robust and efficient way to analyze complex spike train data compared to existing techniques, particularly when information is distributed across multiple time scales.

What Problem Does This Paper Address?

Understanding how the brain encodes information is a fundamental challenge. While the "rate coding" hypothesis suggests that information is primarily conveyed by the average firing rate of neurons, accumulating evidence points to the importance of "temporal coding," where the precise timing of spikes carries crucial information beyond just the rate. However, analyzing spike train data for temporal patterns is notoriously difficult due to the high dimensionality of the data (many spikes, many neurons, over time) and the "curse of dimensionality," which makes it hard to estimate information reliably with limited experimental trials. Existing methods like Principal Component Analysis (PCA) or direct analysis of binned spike trains struggle to capture information spread across different time scales or localized in time. This paper aims to provide a more effective and robust method for extracting this temporal information.

Methods

The core of the proposed method, the Wavelet-Information (WI) approach, involves three main steps:

  1. Wavelet Decomposition: Spike trains are convolved with Haar wavelets. This process decomposes the spike train into coefficients representing local spike patterns at various time scales (from milliseconds to hundreds of milliseconds). This is akin to breaking down a complex signal into its constituent frequencies and time occurrences.
  2. Information-Theoretic Feature Selection: Shannon information theory is used to identify which of these wavelet coefficients carry significant information about the presented stimuli. This is done by calculating the mutual information between each coefficient and the stimuli and using surrogate testing (shuffling trials) to establish statistical significance.
  3. Decoding Performance Assessment: A cross-validated naïve Bayesian decoder is employed to quantify the amount of stimulus information that can be extracted using the selected informative wavelet coefficients. This decoding performance serves as a measure of how well the temporal patterns encode information.

The authors also explore extensions of this method for denoising spike trains and analyzing synchronous firing between multiple neurons.

Datasets

The authors used both simulated and real-world neural data for evaluation:

  • Simulated Data: Generated spike trains mimicking various conditions, including different stimuli, background firing rates, and spike timing jitters. These simulations were designed to test the robustness and performance of the WI method under controlled circumstances.
  • Real Data:
    • Monkey Auditory Cortex: Recordings from primary auditory cortex neurons in response to acoustic stimuli.
  • Rat Barrel Cortex: Recordings from somatosensory cortex neurons in response to whisker stimulation.

Key Insights and Ideas

  • Multiscale Information Encoding: Neural information isn't confined to a single time scale. The WI method's ability to decompose spike trains across multiple wavelet scales is crucial for capturing this distributed information.
  • Wavelets as Efficient Feature Extractors: Wavelets naturally provide a time-localized and multiscale representation of spike trains, effectively reducing dimensionality without losing critical temporal information, unlike methods that rely on fixed bin sizes or global variance.
  • Information-Theoretic Selection is Key: Simply selecting features based on variance (like PCA) isn't optimal for information extraction. Using mutual information to select wavelet coefficients ensures that the chosen features are those that are most predictive of the stimulus.
  • Robustness to Noise and Undersampling: The WI method demonstrates remarkable resilience to background noise, spike timing jitter, and limited trial numbers, making it highly practical for real-world neural data.

Key Results

  • Superior Decoding Performance: The WI method consistently outperformed established methods like PCA, direct binned spike train analysis, and the metric-space (MS) method in decoding stimulus identity from spike trains, both in simulations and real data.
  • Efficiency in Feature Selection: WI required significantly fewer features to achieve near-optimal decoding performance compared to other methods, highlighting its efficiency in dimensionality reduction.
  • Robustness to Parameters: Unlike methods like MS that require careful parameter tuning (e.g., the 'q' value), the WI method is largely parameter-free, making it more user-friendly and less prone to suboptimal settings.
  • Effective Denoising: The WI method was successfully applied to denoise spike trains, leading to clearer visualization of informative temporal patterns and improved stimulus selectivity in analyses like spike-triggered averages (STAs).
  • Multi-Neuron Analysis: The WI framework was extended to analyze synchronous firing between neurons, showing improved performance in extracting information from coordinated neural activity.

Limitations and Future Work

  • Decoding Performance as a Lower Bound: The paper acknowledges that the decoding performance provides a lower bound of the information available in the spike trains. While the WI method is designed to improve this estimation, further development could explore direct information calculations.
  • Computational Cost of MS Method: The MS method was found to be significantly more computationally intensive than the WI method, especially for neurons with higher firing rates.
  • Future Directions: The authors suggest that the WI method could be further developed for direct information calculations, especially for longer response windows, and applied to a wider range of neuroscience problems, including analyzing relationships between spiking activity and local field potentials, and characterizing neural responses in relation to behavior. The availability of a MATLAB implementation is also noted.

Summary updated

Abstract

We present a new method to assess the information carried by temporal patterns in spike trains. The method first performs a wavelet decomposition of the spike trains, then uses Shannon information to select a subset of coefficients carrying information, and finally assesses timing information in terms of decoding performance: the ability to identify the presented stimuli from spike train patterns. We show that the method allows: 1) a robust assessment of the information carried by spike time patterns even when this is distributed across multiple time scales and time points; 2) an effective denoising of the raster plots that improves the estimate of stimulus tuning of spike trains; and 3) an assessment of the information carried by temporally coordinated spikes across neurons. Using simulated data, we demonstrate that the Wavelet-Information (WI) method performs better and is more robust to spike time-jitter, background noise, and sample size than well-established approaches, such as principal component analysis, direct estimates of information from digitized spike trains, or a metric-based method. Furthermore, when applied to real spike trains from monkey auditory cortex and from rat barrel cortex, the WI method allows extracting larger amounts of spike timing information. Importantly, the fact that the WI method incorporates multiple time scales makes it robust to the choice of partly arbitrary parameters such as temporal resolution, response window length, number of response features considered, and the number of available trials. These results highlight the potential of the proposed method for accurate and objective assessments of how spike timing encodes information.
BibTeX
@article{lopesdossantos2014extracting,
  author = {Vítor Lopes‐dos‐Santos and Stefano Panzeri and Christoph Kayser and Mathew E. Diamond and Rodrigo Quian Quiroga},
  title = {{Extracting information in spike time patterns with wavelets and information theory}},
  journal = {Journal of Neurophysiology},
  year = {2014},
  doi = {10.1152/jn.00380.2014},
}
Figure