Neural LSI's Possessing Autonomous Defect Self-repairing Capability

在神经网络中，我们选择自组织映射(SOM)作为目标网络，并对其自主故障修复能力进行了定量评估。特别是，在该项目中，我们提出了缺陷神经元输出任意粘滞值的缺陷模型。通过对模型的分析，证明了如下事实：1)当缺陷神经元的输出大于由所提出的缺陷模型得到的临界粘滞输出时，SOM能够自主修复缺陷神经元。2)新的“临界粘滞输出”准则甚至可以广泛应用于图像压缩和人脸图像识别等实际应用中。此外，在本项目中，我们还评估了进化算法(遗传算法、遗传程序等)的容错性。因为不仅在神经网络中，在其他基于生物信息处理的算法中也可以期待高的容错性。为了评估容错性，我们使用仿真程序和基于可重构LSI(FPGA)构建的原型机进行了故障注入实验。实验结果表明，基于进化算法的硬件对缺陷电路也具有较高的容错性和良好的降级性能。实验结果表明，神经网络LSI和基于进化算法的LSI具有较高的容错性，能够自主修复缺陷电路。

We have chosen the self-organizing map (SOM) as the target network in the neural networks, and evaluated its autonomous fault repairing capability quantitatively. Especially, in the project, we have proposed a defect model in which the defective neurons output arbitrary stuck values. From the analysis of the model, the following facts are shown (proved).1) The SOM can repair the defective neurons autonomously, if the defective neurons'outputs are larger than the critical stuck output, which is derived from the proposed defect model.2) The new criteria "critical stuck output" can be used widely even in the real applications such as image compression and face image recognition.Furthermore, in the project, we have also evaluated fault tolerance of the evolutionary algorithm (genetic algorithms, genetic program etc.) because the high fault tolerance can be also expected not only in the neural networks but also other algorithms based on the biological information processing. In order to evaluate the fault tolerance, we have carried out fault injection experiments using simulation programs and the prototype machine constructed based on reconfigurable LSI (FPGA). From the experimental results, it has been shown that the hardware based on the evolutionary algorithms also has high fault tolerance and graceful degradation against defective circuitsFrom wll those experimental results we have show that neural network LSIs and evolutionary algorithms-based LSIs has high fault tolerance and they can repair the defective circuits autonomously.