A Neural Network Model with interaction between Extraneous Information and Internal Memory

本研究以人类的记忆系统为研究对象，构建了符号模式的增量学习神经网络模型。该网络模型具有层次结构，每一层都有神经元块来识别下层输出模式的局部特征。该网络模型进行自底向上的处理，最终将输入模式识别为局部特征的综合信息。每个神经元块由一个学习触发信号控制，可以独立学习下层的模式。因此，如果输入模式和学习模式之间存在部分差异，网络只能增量地学习它们之间的差异。在处理多感官信息时，网络模型只能学习到尚未学习到的部分信息。在研究中，我们构建了增量学习网络模型。并通过仿真验证了该模型的学习能力。在第一篇文章中，我们使用26种模式作为模拟字母表字符的分布式模式对模型进行了测试。仿真结果表明，该模型可以对所有的模式进行增量学习，并在学习过程中进行识别。然后，我们使用了200个模仿汉字的图案，这些汉字都有“母鸡”和“鸡笼”的结构。每种输入模式都有“母鸡”模式和“土栗”模式的组合，200种输入模式中约有50种“母鸡”模式和50种“土栗”模式。作为模拟的结果，该模型可以学习所有200种模式，只学习“母鸡”和“tsukuri”模式之间的部分差异。我们在1998年3月的IEICE NC98技术组会议上提出了上述结果。

In this study, we focused on the memory system of the human being and constructed a incremental learning neural network model for the symbolic patterns. The network model has a hierarchical structure and each layer has neuron blocks which recognize local features of the lower layer's output pattern. The network model does bottom-up processing and finally recognize the input pattern as the integrated information of the local features. Each neuron block is controlled by a learning trigger signal and can learn the lower layer's patterns independently. Therefore, if there are partial differences between the input pattern and the learned patterns, the network can learn only the differences between them incrementally. In the case of dealing with multi-sense information, the network model can learn only the partial information which the network model have not learned yet. In the study, we constructed the incremental learning network model. And we showed the learning ability of the model by simulations. In the first work, we tested the model by using 26 patterns as the distributed patterns that imitated the alphabet characters. As the result of the simulation, we showed the model could do incremental learning all of the patterns and recognize them after the learning process. And then, we used 200 patterns imitated the kanji characters which had "hen" and "tsukuri" structures. Each input pattern had a combination of "hen" pattern and "tsukuri" pattern, and there were about 50 "hen" patterns and 50 "tsukuri" patterns in the 200 input patterns. As the result of the simulations, the model could learn all of 200 patterns and learn only the partial differences between the "hen" or "tsukuri" patterns. We presented the above results at the IEICE meetings of technical group NC98 (Mar.1998)