适用于神经元状态非等概率分布的神经网络模型及其光学实现

A NEURAL NETWORK MODEL FOR UNEQUALLY DISTRIBUTED NEURON STATES AND ITS OPTICAL IMPLEMENTATION *

针对Ｈｏｐｆｉｅｌｄ模型在存储模式的神经元状态不具备理想的等概率分布时性能下降，以及光学难以实现多灰度阶互连的弱点，提出了一种非对称截值点的截值模型，在易于光学或光电子技术实现的同时，与其他模型相比，存储容量和容噪声能力都有较大提高．同时，提出了光束方向编码方法，并用该方法实现了上述模型，给出了实验结果．

Abstract To avoid the poor performance of the Hopfield model for unequally distributed neuron states and to alleviate the dynamic range constraint of optical system, a clipped model with asymmetric clipping points is proposed. Apart from its easiness for optical implementation, the capacity and capablilty of noisy tolerance are improved greatly, compared with the former clipped model. In addition, a practical encoding method beam direction encoding method is proposed to implement the above clipped model. The preliminary experimental results are given.