Abstract: Interpretable feature selection for classifying context-dependent subnetworks in amygdala
The amygdala is known to be responsible for extracting from external stimuli their social and emotional significance. Contextual changes, which generally take place over longer timescales than stimuli, are crucial to determining the social importance of the stimuli. It is not known how the amygdala encodes social significance. One way the amygdala might encode these persistent states is through context-dependent network dynamics within and between amygdala nuclei. To examine this possibility, we analyzed local field potential (LFP) data recorded from the amygdala of three macaque monkeys presented with alternating blocks of social and non-social tactile stimuli (Gothard Lab, UA). We trained a deep convolutional neural network, LFPNet, to classify spectrograms of baseline (pre-stimulus) LFP activity by stimulus modality (social vs. non-social). To improve efficiency, we use Generalized Eigen Decomposition (GED) for dimensionality reduction. GED provides a set of weights, i.e., a spatial filter, for the recorded channels that maximizes the difference in coactivity patterns between two specified states, e.g., social and non-social baseline LFP signals. Preliminary results show GED dramatically speeds training of LFPNet while maintaining classification accuracy across distinct experimental recording days. Our results indicate that the organization of covarying subnetworks identified by GED are context-dependent, with statistically separable spatial filters that changed between social and non-social blocks. Dimensionality reduction using GED maintains spatiotemporal features crucial to modality classification and provides first steps towards understanding context-dependent network dynamics in amygdala.