Adaptive Sensory-Motor Planning by Humans and Machines

这是授予IRI-9045939的续期。该项目旨在开发神经网络结构，以解释行为个体如何实时成功地适应不断变化的环境施加的限制。这些体系结构能够以这样一种方式进行自动调整，即在不断变化的情况下继续实现高层次的目标和计划。项目包括分析3D空间的表示是如何实时自主学习的。它将解决如何在更大的系统中使用这种空间表示法，该系统能够学习执行熟练的手臂运动序列，例如手写和视觉引导的对象操作。该项目将研究如何在不需要进一步学习的情况下，使用长度和质量可变的工具(如钢笔或挂钩)来执行计划的动作序列。它还将研究工作记忆如何实现对时间顺序代码的可变速率、实时短期记忆存储，例如用于快速语音、打字和眼动序列的代码。这个项目将研究时间规划网络的自组织层次结构如何能够压缩复杂行动序列的控制。这将导致神经网络体系结构如何学习控制接近有价值的目标的方法。最后，这项研究将解决如何将感觉-认知学习系统和认知-运动学习系统的互补设计结合到一个神经架构中，从而实现每种类型设计的好处，而不受各自的限制。所有这些项目都将通过涉及大型行为和神经数据库的计算机模拟来测试其神经结构的生物学相关性。他们还将提供对这些架构的数学和计算分析，以表征具有新实时处理能力的自适应智能机器的电路设计。这个项目将提供对用户在人机交互过程中不同级别的建模的基本见解。

This is a renewal of grant IRI-9045939. This project aims at the development of neural network architectures to explain how behaving individuals adapt successfully in real-time to constraints imposed by changing environments. These architectures are capable of automatically adjusting in such a way that high level goals and plans continue to be realized despite changing circumstances. Projects include an analysis of how representations of 3-D space are autonomously learned in real-time. It will address how such spatial representations can be used in a larger system capable of learning to perform skilled arm movement sequences, such as handwriting and visually-guided object manipulation. The project will examine how planned action sequences can be performed with a tool of variable length and mass, such as a pen or hook, without any further learning. It will also study how a working memory can accomplish variable- rate, real-time short term memory storage of temporal order codes, such as codes for rapid speech, typewriting, and eye movement sequences. This project will examine how a self- organizing hierarchy of temporal planning networks can compress the control of complex action sequences. This will lead into how a neural network architecture can learn to control approach toward a valued goal. Finally, the research will address how complementary designs for sensory-cognitive learning systems and cognitive-motor learning systems can be combined into a neural architecture that achieves the benefits of each type of design, without their individual limitations. All of the projects will test the biological relevance of their neural architectures by computer simulations involving large behavioral and neural data bases. They will also provide mathematical and computational analyses of these architectures to characterize circuit designs for adaptive, intelligent machines with new real-time processing capabilities. This project will provide basic insights into various levels of modeling of a user during human-computer interactions.