Research Initiation: Artificial Neural Network ArchitecturesBased on High-Performance Associative Neural Memories (ANMs)

本项目将研究P.I.在之前的工作中开发的一类新的联想神经记忆（ANM）的能力。这些系统显示出扩展内存容量和可接受的收敛性的希望。特别是，一个借鉴Ho-Kashyap算法的概念进行模式分类的系统，在初步测试中表现优于标准的Hopfield和Kohonen ANM。该项目将进行一系列更全面的测试，包括诊断测试，这些测试将用于指导更强大的ANM的设计。在开发新设计时，pi将特别注意各种算法、多层设计和需要高网络容量的数字乘法等应用中使用的相似性度量。

This project will study the capabilities of a new class of associative neural memories (ANM) developed by the P.I. in prior work. These systems show promise of expanded memory capacity and acceptable convergence. In particular, one system, borrowing concepts from the Ho-Kashyap algorithm for pattern classification, has outperformed the standard Hopfield and Kohonen ANM's in preliminary tests. This project will institute a more comprehensive series of tests, including diagnostic tests which will be used to guide the design of a more powerful ANM. In developing new designs, the P.I. will pay special attention to the similarity measures used in the various algorithms, to multilayer designs, and to applications such as digital multiplication which require high network capacity.