Trans-contextual inference system based only on distributed representations

人们认为，克服经典人工智能局限性的关键是将分布式信息作为模式来表示和处理，而不是将其符号化，并将在特定上下文中学习的知识应用于各种上下文中的新情况。本研究旨在开发一种具有较高的类比推理能力和与人脑推理原理相同的基于模式的推理系统，并取得了以下研究成果。多层神经网络是典型的基于模式的信息处理系统，难以学习依赖于上下文的输入输出关系。为了克服这一困难，我们开发了选择性脱敏方法，利用这种方法，部分神经元件根据上下文进行脱敏。数值实验表明，该模型比传统模型具有更高的学习能力和泛化性能。这是因为两种分布式表示在不使用局部表示的情况下被集成在一起，这极大地扩展了神经网络的潜力。通过将上述方法应用于一种称为轨迹吸引子模型的神经网络，我们构建了一个仅使用分布式表示的推理系统。该系统通过沿大系统中形成的轨迹吸引子的状态转移来推导结论，具有很强的类比推理能力。与现有的推理系统相比，该系统具有许多优点，例如，它可以以一种简单的方式进行非单调推理。为了检验我们推理系统原理的生物学合理性，我们构建了颞叶下皮质和海马区的模型，并将它们与神经生理学数据进行了比较。这一结果与其他生理和心理依据一起，有力地表明选择性脱敏方法实际上用于大脑。

It is considered that keys to overcome the limitations of classical artificial intelligence are to represent and process distributed information as patterns without symbolizing it, and to apply knowledge learned in a particular context to new situations in various contexts. The present study aimed to develop a pattern-based reasoning system that has high ability of analogical reasoning and common principles to that used in the brain, and obtained the following results.1. Multilayer neural networks, which are typical pattern-based information processing systems, have difficulty in learning input-output relation that depends strongly on context. To overcome this difficulty, we developed the selective desensitization method, with which part of neural elements are desensitized depending on the context. We show by numerical experiments that this model has much higher learning capacity and generalization performance than the conventional ones. This is because two kinds of distributed representations are integrated without using local representations, which greatly expands the potential of neural networks.2. By applying the above method to a kind of neural network called a trajectory attractor model, we built a reasoning system using distributed representations alone. This system deduces a conclusion by state transitions along a trajectory attractor formed in a large-scale dynamical system, and has powerful ability of analogical reasoning. We also showed that this system has many advantages over existing reasoning systems, for example, it can make nonmonotonic reasoning in a simple manner.3. To examine biological plausibility of the principles of our reasoning system, we constructed models of inferotemporal cortex and hippocampus and compare them with neurophysiological data. The results, together with other physiological and psychological grounds, strongly suggested that the selective desensitization method is actually used in the brain.