Local learning algorithms for stochastic spiking neural networks

This dissertation focuses on the development of machine learning algorithms for spiking neural networks, with an emphasis on local three-factor learning rules that are in keeping with the constraints imposed by current neuromorphic hardware. Spiking neural networks (SNNs) are an alternative to artificial neural networks (ANNs) that follow a similar graphical structure but use a processing paradigm more closely modeled after the biological brain in an effort to harness its low power processing capability. SNNs use an event based processing scheme which leads to significant power savings when implemented in dedicated neuromorphic hardware such as Intel’s Loihi chip. This work is distinguished by the consideration of stochastic SNNs based on spiking neurons that employ a stochastic spiking process, implementing generalized linear models (GLM) rather than deterministic thresholded spiking. In this framework, the spiking signals are random variables which may be sampled from a distribution defined by the neurons. The spiking signals may be observed or latent variables, with neurons whose outputs are observed termed visible neurons and otherwise termed hidden neurons. This choice provides a strong mathematical basis for maximum likelihood optimization of the network parameters via stochastic gradient descent, avoiding the issue of gradient backpropagation through the discontinuity created by the spiking process. Three machine learning algorithms are developed for stochastic SNNs with a focus on power efficiency, learning efficiency and model adaptability; characteristics that are valuable in resource constrained settings. They are studied in the context of applications where low power learning on the edge is key. All of the learning rules that are derived include only local variables along with a global learning signal, making these algorithms tractable to implementation in current neuromorphic hardware. First, a stochastic SNN that includes only visible neurons, the simplest case for probabilistic optimization, is considered. A policy gradient reinforcement learning (RL) algorithm is developed in which the stochastic SNN defines the policy, or state-action distribution, of an RL agent. Action choices are sampled directly from the policy by interpreting the outputs of the read-out neurons using a first to spike decision rule. This study highlights the power efficiency of the SNN in terms of spike frequency. Next, an online meta-learning framework is proposed with the goal of progressively improving the learning efficiency of an SNN over a stream of tasks. In this setting, SNNs including both hidden and visible neurons are considered, posing a more complex maximum likelihood learning problem that is solved using a variational learning method. The meta-learning rule yields a hyperparameter initialization for SNN models that supports fast adaptation of the model to individualized data on edge devices. Finally, moving away from the supervised learning paradigm, a hybrid adver-sarial training framework for SNNs, termed SpikeGAN, is developed. Rather than optimize for the likelihood of target spike patterns at the SNN outputs, the training is mediated by an auxiliary discriminator that provides a measure of how similar the spiking data is to a target distribution. Because no direct spiking patterns are given, the SNNs considered in adversarial learning include only hidden neurons. A Bayesian adaptation of the SpikeGAN learning rule is developed to broaden the range of temporal data that a single SpikeGAN can estimate. Additionally, the online meta-learning rule is extended to include meta-learning for SpikeGAN, to enable efficient generation of data from sequential data distributions