The Role of Short-Term Synaptic Plasticity in Neural Network Spiking Dynamics and in the Learning of Multiple Distal Rewards

In this thesis, we assess the role of short-term synaptic plasticity in an artificial neuralnetwork constructed to emulate two important brain functions: self-sustained activity andsignal propagation. We employ a widely used short-term synaptic plasticity model (STP)in a symbiotic network, in which two subnetworks with differently tuned STP behaviors areweakly coupled. This enables both self-sustained global network activity, generated by oneof the subnetworks, as well as faithful signal propagation within subcircuits of the othersubnetwork. Finding the parameters for a properly tuned STP network is difficult. Weprovide a theoretical argument for a method which boosts the probability of finding theelusive STP parameters by two orders of magnitude, as demonstrated in tests.We then combine STP with a novel critic-like synaptic learning algorithm, which we callARG-STDP for attenuated-reward-gating of STDP. STDP refers to a commonly used long-term synaptic plasticity model called spike-timing dependent plasticity. With ARG-STDP,we are able to learn multiple distal rewards simultaneously, improving on the previous rewardmodulated STDP (R-STDP) that could learn only a single distal reward. However, we alsoprovide a theoretical upperbound on the number of distal reward that can be learned usingARG-STDP.We also consider the problem of simulating large spiking neural networks. We describean architecture for efficiently simulating such networks. The architecture is suitable forimplementation on a cluster of General Purpose Graphical Processing Units (GPGPU). Novelaspects of the architecture are described and an analysis of its performance is benchmarkedon a GPGPU cluster. With the advent of inexpensive GPGPU cards and compute power,the described architecture offers an affordable and scalable tool for the design, real-timesimulation, and analysis of large scale spiking neural networks.DP