Independent component analyses for quantifying neuronal ensemble interactions

The goal of this study was to compare how multivariate statistical methods for dimension reduction account for correlations between simultaneously rec orded neurons. Here, we describe applications of principal component analys is (PCA) and independent component analysis (ICA) (Cardoso J-F, Souloumiac A. IEE-Proc F 1993;140:362-70; Hyvarinen A, Oja E. Neural Comput 1997;9:148 3-92; Lee TW, Girolami M, Sejnowski TJ. Neural Comp 1999;11:417-41) to neur onal ensemble data. Simulated ensembles of neurons were used to compare how well the methods above could account for correlated neuronal firing. The s imulations showed that 'population vectors' defined by PCA were broadly dis tributed over the neuronal ensembles; thus, PCA was unable to identify inde pendent groupings of neurons that shared common sources of input. By contra st, the ICA methods were all able to identify groupings of neurons that eme rged due to correlated firing. This result suggests that correlated neurona l firing is reflected in higher-order correlations between neurons and not simply in the neurons' covariance. To assess the significance of these meth ods for real neuronal ensembles, we analyzed data from populations of neuro ns recorded in the motor cortex of rats trained to perform a reaction-time task. Scores for PCA and ICA were reconstructed on a bin-by-bin basis for s ingle trials. These data were then used to train an artificial neural netwo rk to discriminate between single trials with either short or long reaction -times. Classifications based on scores from the ICA-based methods were sig nificantly better than those based on PCA. For example, scores for componen ts defined with an ICA-based method, extended ICA (Lee et al., 1999), class ified more trials correctly (80.58 +/- 1.25%) than PCA (73.14 +/- 0.84%) fo r an ensemble of 26 neurons recorded in the motor cortex (ANOVA: P < 0.005) . This result suggests that behaviorally relevant information is represente d in correlated neuronal firing and can be best detected when higher-order correlations between neurons are taken into account. (C) 1999 Elsevier Scie nce B.V. All rights reserved.