Locally-connected VLSI neural networks for invariant pattern recognition

Recognizing the components that make biological systems so effective for certain tasks and functionally capturing what is required to best take advantage of the medium of implementation are the keys to realizing useful biologically inspired artificial systems. A compact set of novel integrated circuits was designed based on the Baek biomorphic spiking neuron model that includes both the spiking behavior at the axon hillock and the synapto-dendritic processing found in biological neurons. A 12 x 12 locally-connected array of these neuron and synapse circuits was fabricated into a pulse-coupled neural network (PCNN) on a custom microchip to explore the network's ability to generate pattern feature vectors to be used for pattern recognition. Testing and analysis of the fabricated chip inspired a new and more efficient locally-connected neural network (LCNN) architecture capable of generating feature vectors that capture local, global, and relational characteristics of an input pattern. Computer simulations of this new architecture indicate an ability to perform pattern recognition invariant to rotation and translation when combined with a simple classifier. The system achieved a recognition rate of almost 97% when applied to optical character recognition, and almost 99% when applied to fingerprints. Simulations of analog integrated circuit designs to implement the LCNN show the new architecture has faster convergence, lower power, and increased integration density over the PCNN. The hardware implementation of the fully parallel architecture will provide a six order-of-magnitude advantage in both speed and energy over computing on conventional microprocessors. The LCNN architecture and hardware implementation could pave the way to the integration of sensor and processing, analogous to visual preprocessing in biological systems, that enables image acquisition and pattern recognition on a single chip.