BIC: Bio-inspired Information Processing Using Hybrid Nanodevice Arrays

BIC：使用混合纳米器件阵列的仿生信息处理

• 批准号: 0432116
• 负责人: Konstantin Likharev
• 金额: $ 15万
• 依托单位: SUNY at Stony Brook
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2004
• 资助国家: 美国
• 起止时间: 2004-08-15 至 2006-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0432116&HistoricalAwards=false
• 关键词: BIC; Bio; inspired; Information; Processing

(i) Technical Description: Preliminary studies indicate that hybrid semiconductor/molecular ("CMOL") circuits may serve as a basis for neuromorphic network arrays ("CrossNets") capable of advanced mixed-signal information processing with unprecedented density (beyond 1012 active devices per cm2) and performance (up to 1020 elementary operations per second per cm2) at acceptable power consumption (below 100 W/cm2). The objective of this project is to the key issue of the CMOL CrossNets: the development of effective techniques of their training (including the error backpropagation and global reinforcement), and the demonstration of the latter technique of such an important task as modeling attention. (ii) Non-technical Explanation: Preliminary estimates indicate that array circuits, based on both semiconductor (CMOS) devices and nanoelectronic (e.g., molecular) devices and using bio-inspired architectures, may provide advanced information processing with much higher performance than the usual semiconductor Boolean logic circuits. However, in order to enable such processing, the existing methods of neural network training should be modified to accommodate the limitations imposed by the hybrid array hardware. This project focuses on the development of two most prospective training techniques, and their demonstration for modeling such a key function of brain as attention.

(i)技术描述：初步研究表明，混合半导体/分子（“CMOL”）电路可以作为神经形态网络阵列（“crosnet”）的基础，能够以前所未有的密度（每平方厘米超过1012个有源器件）和性能（每平方厘米每秒高达1020个基本运算）进行高级混合信号信息处理，功耗可接受（低于100 W/平方厘米）。本项目的目标是解决CMOL交叉网的关键问题：开发有效的交叉网训练技术（包括误差反向传播和全局强化），并演示后者技术在建模注意力等重要任务中的应用。（ii）非技术解释：初步估计表明，基于半导体（CMOS）器件和纳米电子（如分子）器件并使用生物启发架构的阵列电路可以提供比通常的半导体布尔逻辑电路更高性能的高级信息处理。然而，为了实现这种处理，现有的神经网络训练方法应该进行修改，以适应混合阵列硬件所施加的限制。本项目的重点是两种最有前景的训练技术的发展，以及它们对大脑的关键功能（如注意力）建模的演示。