Complex Neural Computation with Simple Digital Neurons.

摘要

The desire to understand, simulate, and capture the computational capability of the brain is not a new one. However, in recent years, many advances have been made towards building better models of neurons and cortical networks. Furthermore, a number of high profile projects have proposed, designed, and fabricated neuromorphic substrates inspired by the structure, organization, and behavior of biological brains.;This dissertation explores both the software and hardware elements of these neuromorphic systems. On the software side, this dissertation begins with an exploration of the leaky integrate-and-fire (LIF) spiking neuron, and demonstrates that a network composed of simple LIF neurons is capable of simple object recognition and motion detecting tasks. Furthermore, a number of complex neuronal behaviors which significantly extend the computational power of the LIF neuron are identified. This dissertation proposes that an extended LIF neuron model can be used to construct a large scale functional model of the visual cortex with metastable attractor dynamics. This hierarchical metastable attractor is capable of invariant object recognition, image reconstruction, working memory tasks, and demonstrates functional integration across multiple modeled regions.;On the hardware side, this dissertation investigates the challenges associated with neuromorphic hardware in the context of IBM's Neurosynaptic Core. This neuromorphic substrate, composed of simple digital neurons, highlights the neuromorphic semantic gap that exists between software models such as the ones described in this dissertation, and the hardware on which they will be deployed. This dissertation demonstrates how this semantic gap can be effectively bridged, and proposes a number of automated techniques for deploying large scale cortical models on the Neurosynaptic Core hardware.