Symbol Processing System Modeled after Brains

我们研究了一种新的神经网络学习框架，该框架由多个强化学习代理组成，其中存在多个合法的候选模块。我们发明了一种促进强化学习代理之间竞争学习的机制，并通过计算机模拟确定了其有效性。我们进一步研究了两种不同类型的递归神经网络的另一种语法习得。其中一种类型监视另一种类型，并根据被监视的观察结果修改自身。我们发现，这些网络能够学习语法类别，并且对大量语法类别具有很强的抵抗力。我们以Penn TreeBank中的ATIS语料库为语料库，以ILP为基本学习范式，进行了shift-reduce解析器的获取实验。为了减轻现有学习方法的缺点（高执行时间成本和内存需求），我们采用语法类别作为学习单元。我们发明了新的方法来根据语法类别生成合理化的否定例子，并通过调查这些例子实际上是错误分类的地方来重新学习否定例子。我们证实，准确度提高到略低于90%，有限状态自动机有足够的能力来表示大脑中的知识，但众所周知，它们太通用而无法成功学习。因此，我们对如何近似地学习和交流它们进行了研究。我们已经应用了强化学习方法，并发现它们是合格的。我们发明了一种方法来准备大量的语法类别，并尝试在交流中使用它们，并选择最好的。我们在递归神经网络中实现了该方法，并进行了数值模拟。从某种意义上说，结果是有希望的，原始的语法类别结构被重构，只关注训练错误（不关注泛化能力）。

We have investigated a new framework of neural network learning that is composed of multiple reinforcement learning agents among which there exist multiple legitimate candidate modules. We have invented a mechanism that facilitates competitive learning among reinforcement learning agents and ascertained its validity by computer simulations.We further investigated another type of grammar acquisition by recurrent neural networks of two different types. One type of them monitors the other type and modifies itself based on the monitored observation. We found that the networks are able to learn grammatical categories and are robust against their legion.We conducted experiments of acquisition of shift-reduce parsers in which ATIS corpus in Penn TreeBank is the corpus and ILP is the fundamental learning paradigm. To alleviate drawbacks (high cost of execution time and memory requirements) of the existing learning methods, we employed grammatical categories as learning units. We invented new methods to generate rationalized negative examples based on grammatical categories and to relearn the negative examples by investigating where those examples are in fact miss-classified. We confirmed that the accuracy improved to a bit less than 90%,Finite state automata are capable enough to represent knowledge in brain but it is well-known that they are too versatile to be successfully learned. Therefore we made research on methods to approximately learn and communicate them. We have applied reinforcement learning methods and found them to be eligible. We have invented a method to prepare a large number of grammatical categories, to try to use them in communication, and to select best ones. We implemented the method in recurrent neural networks, and conducted numerical simulations. The results are promising in a sense that the original grammatical category structure is reconstructed with paying attention only to training errors (not to generalization capabilities).