Filter Bank Common Spatial Pattern (FBCSP) is the winner of BCI Competition IV Datasets 2a and 2b. I implemented this algorithm and applied it to a larger EEG ‘motor imagery’ dataset from Kaya et al., 2018. My results outperform the SVM-based approach taken by Kaya et al. 
I discovered subpopulations of inhibitory neurons that perform distinct computations and provide flexible coding across shifts in input statistics
Separate populations of inhibittory neurons perform input gain (left) and response gain (right). These populations are differentially recruited based on feature sparsity, and mutually inhibit one another. Therefore the network alters its normalization scheme based on input statistics.