Learning and Adaptation in the Primate Oculomotor System: A Neuromorphic Analog VLSI Model

9720353尼布尔学习和适应不断变化的环境的能力是生活本身的基本特征之一。拟议工作的目的是开发、实施和测试仿照其生物对应机制的适应机制。该项目的中心是开发自适应的“神经形态”VLSI(超大规模集成)设备。神经形态设备类似于常见的计算机芯片，但它们的设计和操作方式略有不同(模拟变量而不是数字，亚阈值电压而不是阈值以上电压，并行处理而不是串行处理)，这赋予了它们比标准计算机芯片更类似于神经细胞的特性。虽然神经形态电路的设计和制造已经有大约十年的时间，但这项工作是第一次系统地将学习和适应纳入这项技术。这是通过利用最近开发的设计技术(‘浮动门’)来实现的，该技术首次允许在与动物和人类遇到的保持机制的寿命类似的时间段(年)内存储获得值。此外，虽然以前的大部分工作都是关于感官信息的处理(很好的例子是“硅视网膜”和“硅耳蜗体”)，但这个项目也包括运动控制，从而关闭了从感觉输入到运动输出的行为环路，这又一次影响了感觉输入。自适应神经形态设备被集成在灵长类动物(猴子和人类)眼动控制的现实模型中，这是最被理解的生物感觉运动系统之一。在猴子和人类身上研究了不同的适应和学习机制，并通过与实验数据的比较来评估自适应控制系统的性能。眼球运动硬件模型的开发为了解一个重要的感觉运动控制系统的功能提供了重要的见解。了解现实世界中人类(和灵长类)的眼睛控制，以及它如何适应不断变化的情况，可能会导致开发实用的工具，例如，用于目标检测和追踪。然而，更重要的是，合成一种复杂的行为将帮助我们在一定程度上理解这种行为模式和其他行为模式，这是单靠分析方法不可能实现的。

9720353 Niebur The ability to learn and to adapt to a changing environment is one of the fundamental characteristics of life itself. The purpose of the proposed work is to develop, implement and test adaptation mechanisms which are modeled after their biological counterparts. The centerpoint of the project is the development of adaptive "neuromorphic" VLSI (Very-Large-Scale-Integrated) devices. Neuromorphic devices are similar to common computer chips but they are designed and operated in somewhat different regimes (analog variables instead of digital, subthreshold voltages instead of suprathreshold, and parallel processing instead of serial) which gives them properties much more similar to those of nerve cells than can be claimed of standard computer chips. While neuromorphic circuits have been designed and fabricated for about a decade, this work is one of the first to systematically incorporate learning and adaptation into that technology. This is accomplished by making use of recently developed design techniques (''floating gates'') which for the first time allow storage of learned values for periods of time (years) similar to the lifetimes of retention mechanisms encountered in animals and humans. Furthermore, while most of the previous work was concerned with the processing of sensory information (good examples being the "silicon retina" and "silicon cochlea"), this project also includes motor control and thus closes the behavioral loop, from sensory input to motor output which then again influences sensory input. The adaptive neuromorphic devices are integrated in a realistic model of the eye-movement control of primates (monkeys and humans), which is one of the best understood biological sensorimotor systems. Different mechanisms of adaptation and learning have been studied in monkeys and humans and the performance of the adaptive control system is evaluated by comparing with experimental data. The development of a hardware model for eye movements promises significant insights into the function of an important sensorimotor control system. Understanding human (and primate) eye control in the real world and how it adapts to changing situations may lead to the development of practical tools, for instance, for target detection and pursuit. More importantly, however, synthesizing a complex behavior will help us understand this and other behavioral patterns to a degree which is impossible to achieve with analytical methods alone.