This dissertation is concerned with biologically inspired artificial neural networks, which offer spatio-temporal signal-processing capabilities that conventional neural nets cannot account for. Most conventional neural nets use sigmoidal input-to-output transfer functions, based on the mean firing frequency of neurons, but the time scale of the transfer function is too coarse to describe the intricate dynamics of nerve impulses in biological neural networks. By incorporating essential features of the biological neuron such as action potential generation and dendritic-tree processing, we capture the dynamics that emerge from interactions of action potentials (nerve impulses) in what is called pulse-coupled networks.;In achieving such a goal, we employed a biology-like model neuron called biomorphic spiking neuron. The model neuron can exhibit in its firing modalities various functional complexities under periodic stimulation, such as phase-locking and chaos, and bifurcation between the two states. The model neuron also contains a linear dendritic tree, which functions as spatio-temporal integrator of action potential inputs. The dendritic tree plays a vital role in neuronal signal processing in that it can generate a periodic signal when it receives on its synapses correlated spike trains supposedly from phase-locked neurons. Our motivation to drive the biomorphic spiking neuron with a periodic signal comes from this concept of the periodic signal generation from the dendritic processing.;In exploring applications, the biomorphic neuron showed great potential in sensitive signal detection and invariant feature extraction. Especially, the invariant feature extraction from canonical patterns showed robust invariance under translation, scale, and rotation. Since invariant feature extraction is central in human vision, the extraction method can be applied to many different classes of images. We extended the method to handwritten signatures and segmented images of model military objects. The results show that biomorphic spiking neuron models offer advantages over conventional sigmoidal neuron models.;We also address the issue of hardware implementation for speed and efficiency, which is an important subject in the study of neural networks. We provide a novel hardware implementation scheme of an optoelectronic biomorphic spiking neural network, which consists of optically coupled biomorphic spiking neurons with optically programmable dendritic trees realized in electron trapping materials (ETMs).;This dissertation presents evidence that biomorphic spiking neural networks incorporating dendritic-tree processing offer a new computing paradigm based on spatio-temporal interaction of nerve impulses and demonstrate their potential in sensitive signal detection of sensory information and invariant feature extraction for automated recognition systems.
展开▼
