In simulations of recurrent neuronal networks, spike timing dependent plasticity (STDP) is often found to cause a self enforcing cycle of increasing connection weights and spiking rates, known as 'runaway excitation'. This network transformation might terminate in a synchronized high frequency activity or biological implausible bimodal spiking rate and weight distributions.
It could be demonstrated that a simple STDP model with an asymmetric weight dependence of the potentiation and depression term can support the formation, and within certain conditions, the maintenance, of a biologically realistic connection weight distribution in a recurrent spiking network. The resulting network activity is characterized by a low spiking rate and a small pairwise spike time correlation coefficient.
The stability of these network properties depends on the precise fine tuning of one of the STDP parameter, balancing the average size of STDP potentiation and depression weight changes. The point of balance is effectively depending on the spike timing correlation within the network. The dependence of the network stability on a stable spike timing correlation can be avoided through the addition of a homeostatic plasticity (HP) mechanism that is controlling the spiking rates of the excitatory neurons through a negative feedback process on the afferent excitatory connection weights.
If the STDP process is faster than the HP process, the spiking rate of a subset of neurons can escape the rate homeostasis, causing a network disruption with some of the neurons engaging in high frequency activity while the other neurons are silenced. This behavior was common in networks with 4000 excitatory neurons but absent (within simulation duration) in networks with 10000 excitatory neurons.
The talk is open to the public. Guests are cordially invited!