A0701
Title: Spectral causation entropy and its amortized neural estimator
Authors: Paolo Victor Redondo - King Abdullah University of Science and Technology (Saudi Arabia) [presenting]
Raphael Huser - King Abdullah University of Science and Technology (Saudi Arabia)
Hernando Ombao - King Abdullah University of Science and Technology (KAUST) (Saudi Arabia)
Abstract: Given several nodes in a network, effective connectivity describes the causal relationship between processes recorded at two specific nodes. However, existing methods construct such connectivity mapping by considering pairwise analysis, which neglects the contribution of other components in the network and is unable to differentiate direct from indirect causal structures. A solution is offered by the metric called causation entropy, an information-theoretic causal measure that quantifies the magnitude and direction of information flow between two processes after taking into account all other processes in the network. To associate a derived network with well-established findings in cognitive science, a new spectral causal measure is developed, the spectral causation entropy (SCE), that measures the direct causal impact in the frequency domain. By combining the vine copula theory with the amortized neural inference approach, an estimator is developed for SCE, which can provide both an estimate and confidence bounds for the metric in milliseconds after training the neural network. Based on SCE, summarizing significant direct information flow from all node pairs in the network results in the derivation of the spectral effective connectivity graphs. Lastly, the performance of the proposed measure is demonstrated through numerical experiments, and interesting findings on the analysis of electroencephalogram (EEG) recordings linked to a virtual driving task are reported.
