Hypothetical reasoning and brainware

摘要

An enormous variety of complex problems requiring tentative interpretation of real-world situations or conditions, such as the visual recognition of natural scenes, as the basis for some action cannot be properly managed without the effective use of acquired knowledge to reduce the complexity dramatically. There is much evidence that the brain solves this type of problem by creating an internal hypothesis from its store of acquired knowledge. It then reformulates the task as essentially one of simply comparing this internally generated hypothesis with the objective sensory reality signaled to it by the sensors (such as the eye). Despite a rich body of research on hypothetical reasoning in the field of AI, this method of generating and verifying a hypothesis bridging signal and symbol levels has never been demonstrated. A neural control architecture that uses this problem-solving methodology is proposed in this paper. Multilayer neural network architecture, modeled after the essential features of the cortical structure and function, is used to simulate hypothetical reasoning and in particular to demonstrate its role in the marvelous performance of visual recognition. With this architecture it is possible not only to control the logical reasoning steps to ensure a rapid convergence in the decision-making processes but also to set the constraints for the self-organization of knowledge required in creating the initial internal hypothesis.