Indirect Training Algorithms for Spiking Neural Networks based on Spiking Timing Dependent Plasticity and Their Applications

Abstract

<p>Spiking neural networks have been used to investigate the mechanisms of processing</p><p>in biological neural circuits or to propose hypotheses that can be tested in exper-</p><p>iments. Because of their biological plausibility and event-based information trans-</p><p>mission, Spiking Neural Networks (SNNs) have suggested as alternatives to Articial</p><p>Neural Networks for pattern recognition, classication and function approximation</p><p>problems with fewer neurons. In machine learning, SNNs has been shown to be able</p><p>to solve pattern and robotic control. For SNNs to be used for such problems, they</p><p>must incorporate some mechanism for learning. Current methods to train SNNs use</p><p>learning algorithms which adjust the synaptic weights according to an update rule.</p><p>In most cases the weights are modied directly. In potential applications such as</p><p>driving plasticity in neural culture (in-vitro) and training neuromorphic chips the</p><p>directly manipulation of synaptic weights is not possible. Therefore, indirect algo-</p><p>rithms, which cause the SNNs to learn based on some biological learning mechanisms</p><p>using stimulation of neurons oer signicant advantages over the existing algorithm</p><p>for these real world applications.</p><p>Indirect algorithms train the neural network by using external stimuli to modulate</p><p>the synaptic strengths of a neural network according to synapses intrinsic mechanisms</p><p>for plasticity. The training algorithms have been demonstrated in both Integrate and</p><p>Fire neurons and more biologically realistic neural networks. In this thesis, four indi-</p><p>rect methods to drive the synaptic weights to its desired value in a network through Spike Time Dependent Plasticity (STDP) are developed: Indirect Perturbation, In-</p><p>direct Stochastic Gradient, Indirect ReSuMe, and Indirect Training with Supervised</p><p>Teaching Signals. These algorithms are used to solve the temporal and spatial input-</p><p>output mapping problem using temporal coding. The other type of problem is to</p><p>mapping input output ring rates using rate coding.</p><p>To test the algorithms, SNNs are used to control both virtual and real world</p><p>robots. For the real world robots with SNNs, known and Neurorobots, two types</p><p>of robot localization techniques are used: Optitrack, using ceiling mounted cameras</p><p>and onboard markers, and embedded cameras. Both small and large SNNs with</p><p>biologically realistic neurons are used to drive the neurorobots are modeled with</p><p>input coming from Optitrack or the cameras with GPU accelerated SNN simulator.</p><p>The results show that the indirect perturbation and indirect stochastic gradient</p><p>algorithms can train an SNN to control the robot to nd targets and avoid obstacles</p><p>even in the presence of sensor noise. The results also show that indirect training with</p><p>supervised training signals algorithm can train a feedforward network with 1000s of</p><p>neurons to process and output the correct movement commands to localize a target</p><p>using from real time images captured from an embedded camera. Finally. an indirect</p><p>version of the Remote Supervised Method (ReSuMe) algorithm was developed using</p><p>a more biologically realistic form of Spike-Timing Dependent Plasticity to produce a</p><p>specic temporal pattern of spiking from a group of neurons. The indirect algorithms</p><p>developed in this thesis may eventually allow the ability to train in vitro and in</p><p>vivo biological circuits to perform specic tasks using patterns of electrical or light</p><p>stimulation.</p>