Fast gradient-free activation maximization for neurons in spiking neural networks

Nikita Pospelov, Andrei Chertkov, Maxim Beketov, Ivan V. Oseledets, Konstantin Anokhin

Abstract: Highlights•We propose a new activation maximization method based on Tensor Train decomposition.•We apply our method to obtain optimal stimuli of neurons in a spiking CNN.•This is the first time that gradient-free AM methods have been applied to SNNs.•We study the computed optimal stimuli layer-wise throughout the network training.•We correlate model performance with the formation of highly selective neurons.•We find selectivity patterns similar to the ones observed in in vivo brain studies.•Potential applications of the proposed method to biological systems are discussed.