Implement of Hopfield Neural Network using Grating Approach and Its Applications

• Main Authors: Jia Yang Sheu, 許家暘
• Other Authors: Rong-Chin Lo
• Format: Others
• Language: en_US
• Published: 2001
• Online Access: http://ndltd.ncl.edu.tw/handle/04890353183967486230

碩士 === 國立臺北科技大學 === 光電技術研究所 === 89 === Optical Hopfield Neural Network is one of associative-memory system. It can recall entire pattern from partial information and has error correction capability. However in general neural network, full interconnects of an NxN input signal array to an NxN output signal array require N4 weighted interconnects. To implement the neural networks, full connections between neurons facility and high-speed computing always are both critical problems. Up to today, the improved methods for this study are still rather few. Early stage, the fundamental method to connect the neurons using lenslet array can achieve optical matrix-vector multiplier, but connection density is low. Later stage, holography is invented, which can store plenty of information by interference. It is cheaper than lens set and is often utilized in optical neural network. Although holography is a good tool to optical system, but it also has some drawbacks, such as we cannot easy to input any pattern because the holography needs to be repictured. On the contrary, the proposed system can perform dynamic processes in real time and is more convenient to input any pattern than holography. In this paper we propose an optical grating approach to improving the Hopfield connection better, and combining the liquid crystal display (LCD) and grating plate to realize the Hopfield model by diffractive theory. In the optical Hopfield model, the LCD acts as input device to input the stimulus pattern and is directly controlled by the computer. The grating array designed with photoresist plate is used to control the diffracted direction to focus the input light of LCD from different positions into the same block of output plane, i.e., CCD camera. The process acts like input summation of a neuron. Several experimental results are also included to demonstrate the system is high-speed computing and feasible recalling capability.