Constructive spiking neural networks for simulations of neuroplasticity

Abstract

Artificial neural networks are important tools in machine learning and neuroscience; however, a difficult step in their implementation is the selection of the neural network size and structure. This thesis develops fundamental theory on algorithms for constructing neurons in spiking neural networks and simulations of neuroplasticity. This theory is applied in the development of a constructive algorithm based on spike-timing- dependent plasticity (STDP) that achieves continual one-shot learning of hidden spike patterns through neuron construction. The theoretical developments in this thesis begin with the proposal of a set of definitions of the fundamental components of constructive neural networks. Disagreement in terminology across the literature and a lack of clear definitions and requirements for constructive neural networks is a factor in the poor visibility and fragmentation of research. The proposed definitions are used as the basis for a generalised methodology for decomposing constructive neural networks into components to perform comparisons, design and analysis. Spiking neuron models are uncommon in constructive neural network literature; however, spiking neurons are common in simulated studies in neuroscience. Spike- timing-dependent construction is proposed as a distinct class of constructive algorithm for spiking neural networks. Past algorithms that perform spike-timing-dependent construction are decomposed into defined components for a detailed critical comparison and found to have limited applicability in simulations of biological neural networks. This thesis develops concepts and principles for designing constructive algorithms that are compatible with simulations of biological neural networks. Simulations often have orders of magnitude fewer neurons than related biological neural systems; there- fore, the neurons in a simulation may be assumed to be a selection or subset of a larger neural system with many neurons not simulated. Neuron construction and pruning may therefore be reinterpreted as the transfer of neurons between sets of simulated neurons and hypothetical neurons in the neural system. Constructive algorithms with a functional equivalence to transferring neurons between sets allow simulated neural networks to maintain biological plausibility while changing size. The components of a novel constructive algorithm are incrementally developed from the principles for biological plausibility. First, processes for calculating new synapse weights from observed simulation activity and estimates of past STDP are developed and analysed. Second, a method for predicting postsynaptic spike times for synapse weight calculations through the simulation of a proxy for hypothetical neurons is developed. Finally, spike-dependent conditions for neuron construction and pruning are developed and the processes are combined in a constructive algorithm for simulations of STDP. Repeating hidden spike patterns can be detected by neurons tuned through STDP; this result is reproduced in STDP simulations with neuron construction. Tuned neurons become unresponsive to other activity, preventing detuning but also preventing neurons from learning new spike patterns. Continual learning is demonstrated through neuron construction with immediate detection of new spike patterns from one-shot predictions of STDP convergence. Future research may investigate applications of the developed constructive algorithm in neuroscience and machine learning. The developed theory on constructive neural networks and concepts of selective simulation of neurons also provide new directions for future research.