Training oscillatory dynamics with spike-timing-dependent plasticity in a computer model of neocortex

Neuronal networks are complex, adaptive systems that typically display oscillatory dynamics. The extent to which these dynamics can be shaped by training remains unknown. We explored this dynamical training in a computer model of 6-layered sensory neocortex with 470 excitatory (E) and inhibitory (I) cells. AMPA, NMDA, and GABAA synapses were provided with Poisson input to provide baseline activation in the network. The learning rule employed spike-timing-dependent plasticity (STDP) at all AMPA synapses. We trained with a 1-16 Hz thalamic afferent signal to E4 cells (layer 4 E cells). At baseline, the power spectrum of the network activity showed oscillations with a low-amplitude peak near 6 Hz. Plasticity in the absence of a training signal (white noise input) attenuated the network response, due to the potentiation of E-to-I synapses. Plasticity coupled with an 8 Hz training signal enhanced the network´s oscillations and shifted the peak to ~20 Hz. This was due to increased synaptic connection strengths between E cells caused by the near-synchronous firing of E4 cells. Plasticity coupled with a 16 Hz training signal shifted the network towards epilepsy, with high-amplitude 8 Hz oscillations and synchronous firing across all layers. The shift into epilepsy was caused by further enhancement of E-to-E synapses. In summary, our simulations demonstrate the feasibility of using plasticity and neuroprosthetic input signals to train a neuronal network´s oscillatory dynamics. We predict that in order for learning in the brain to avoid transition to epilepsy, homeostatic control mechanisms must balance learning at E-to-E and E-to-I synapses.