The complex processes of learning and memory in vivo often involve simultaneous alterations in synaptic strength and intrinsic excitability of the constituent neurons. However, such synergy is rarely explicitly manifested in the bottom-up training of spiking neural models, partially because of conceptual and computational difficulties. Here we incorporate the combined plasticity mechanisms into a spike response model (SRM), by which postsynaptic responses therein may deviate from stereotyped properties and evolve into arbitrary, yet self-consistently determined kernels for adaptation to imposed activity. We develop a general framework for reconstruction of adaptive SRMs as a neuroinspired generative model of event sequence data. It provides a descriptive approach to generate the observed timing of spiking activity of a set of interconnected neurons with nonparametric response kernels determined by event timing data. We further illustrate the potential of this neurally inspired model by two examples, time-series extreme event forecasting and electroencephalogram classification. Our approach provides a new conceptually simple but effective generative model for recognizing and exploiting event timing patterns in a broad variety of applications.
