The Supervised Learning Rules of the Pulsed Neuron Model

提出了一种利用脉冲神经元模型(PNM)来处理以脉冲形式表示的信息的听觉神经网络模型。在本研究以前的进展中，PNM中使用的参数没有自动配置。因此，很难将该模型应用到实际应用中。这一局限性成为了为脉冲神经元模型构建通用算法的动机。在本研究的第一年，我们提出了两种适合基于PNM实现的新算法。第一种是有监督算法，用于学习脉冲之间的特定相位差。第二种是无监督模式分类算法。通过实现这些算法，PNM具有了非线性映射的能力，使其适合于实际领域的应用。第二年，对有监督算法在声源定位问题和声源识别问题上的性能进行了评估。实验是在真实声音数据库的条件下进行的。在这些研究中，我们还试图在一个带有各种声音的环境中对该系统进行评估。针对这种情况，在系统中增加了一个时差提取模块作为预处理部分，对输入信号的声音进行分类。实验结果表明，该系统具有从特定方向识别特定声音的能力。这项研究的一部分还致力于PNM硬件实施的成本评估。提出了一种将PNM实现为逻辑电路的有效方法，并进行了实验，取得了良好的效果。这项研究的进展于1999年12月、2001年3月在IEICE神经计算技术小组研讨会、1999年国际神经信息处理会议(1999年11月)和2000年国际神经网络联席会议(2000年7月)上介绍。

We propose an Auditory Neural Network Model using the Pulsed Neuron Model (PNM) to deal with the information which is presented as pulse. In the previous progress of this study, the parameters used in PNM were not configured automatically. Therefore, it was difficult to implement the model into real applications. This limitation became the motivation to build a general-purpose algorithms dedicated for the Pulsed Neuron Model. In the first year of this study, we proposed 2 new algorithms which were suitable for PNM based implementation. The first one was supervised algorithm to learn the specific phase-difference between the pulses. The second one was an unsupervised algorithm for pattern classification. By implementing these algorithms, PNM had the capability of non-linear mapping to make it suitable for real-world domain application. In the second year, the performance of the supervised algorithm was evaluated on sound source localization problem and sound source recognition problem. The experiments were performed under the condition which was using real-life sound database. In these study, we also attempted to evaluate the system in an environment which comes with various sounds. To deal with this condition, a time-difference extraction module was added to the system as preprocessing part to classify the sound of the input signal. The experiment results showed that the system had the capability to recognize specific sound from specific direction. A part of this study was also dedicated on the cost evaluation for the hardware implementation of PNM. We proposed efficient method to implement PNM as logic circuit and the experiments showed promised result. The progress of this study was presented in IEICE Neurocomputing-Technical Group Workshop on December 1999, March 2001, International Conference On Neural Information Processing '99 (November 1999) and International Joint Conference on Neural Network 2000 (July 2000).