Covariance shrinkage for autocorrelated data

Authors: Daniel Bartz, Klaus-Robert Müller

NeurIPS 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We propose an alternative estimator, which is (1) unbiased, (2) less sensitive to hyperparameter choice and (3) yields superior performance in simulations on toy data and on a real world data set from an EEG-based Brain-Computer-Interfacing experiment.
Researcher Affiliation Academia Daniel Bartz Department of Computer Science TU Berlin, Berlin, Germany daniel.bartz@tu-berlin.de Klaus-Robert M uller TU Berlin, Berlin, Germany Korea University, Korea, Seoul klaus-robert.mueller@tu-berlin.de
Pseudocode No The paper presents mathematical formulas and descriptions of estimators but does not include structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any specific repository links or explicit statements about the public release of the source code for the methodology described.
Open Datasets Yes As an example of autocorrelated data we reanalyzed a data set from a motor imagery experiment. In the experiment, brain activity for two different imagined movements was measured via EEG (p = 55 channels, 80 subjects, 150 trials per subject, each trial with ntrial = 390 measurements [BSH+10]).
Dataset Splits No The paper analyzes performance based on the 'number of trials per class' (training data size) and mentions cross-validation as a baseline method. However, it does not provide specific dataset split information (exact percentages, sample counts, or predefined split citations) needed to reproduce the data partitioning for their own experiments.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments or simulations.
Software Dependencies No The paper does not provide specific ancillary software details, such as library or solver names with version numbers, needed to replicate the experiment.
Experiment Setup Yes Our simulations are based on those in [San08]: We average over R = 50 multivariate Gaussian AR(1) models xt = A xt 1 + ϵt, with parameter matrix4 A = ψAC I , with ψno AC = 0, ψlow AC = 0.7, and ψhigh AC = 0.95 (see Figure 1). The innovations ϵit are Gaussian with variances σ2 i drawn from a log-normal distribution with mean µ = 1 and scale parameter σ = 0.5. For each model, we generate K = 50 data sets to calculate the std. deviations of the estimators and to obtain an approximation of p 2 P ij Var (Sij). ... The number of observations is fixed at n = 250 and the lag parameter b chosen by hand such that the whole autocorrelation is covered5: bno AC = 10, blow AC = 20 and bhigh AC = 90. ... for a motor imagery experiment. In the experiment, brain activity for two different imagined movements was measured via EEG (p = 55 channels, 80 subjects, 150 trials per subject, each trial with ntrial = 390 measurements [BSH+10]). ... The frequency band was optimized for each subject and from the class-wise covariance matrices, 1-3 filters per class were extracted by Common Spatial Patterns (CSP), adaptively chosen by a heuristic... We trained Linear Discriminant Analysis on log-variance features. ... an optimized time lag (b = 75) ... a too large time lag (b = 300).