High Performance Associative Memory: Practical Generalizations of the Hamming Net

点击翻译按钮获取中文摘要

The design of high performance associative memory has long been a goal of artificial neural net researchers. A practical low-cost device which could store and retrieve information similar to the human `content addressable` memory would be recognition, data base, control systems, and robotics. Indeed, associative memory is seen by most experts as a crucial component for the development of any truly intelligent and brain-like system. The purpose of this project is to design and implement a practical and low cost associative memory with following properties: 1. High capacity. 2. High amount of error correction and tolerance to noisy or partial inputs. 3. No spurious memories and no limit cycles. 4. Provision for a `no decision` or ground state. 5. Fast memory retrievals. 6. Ability to operate in an autoassociative or heteroassociative mode. 7. Practical and low cost hardware implementation. The approach will be to generalize the operation of the Hamming memory in order to design associative memory models which retain all the high performance properties of the Hamming net, but allow for a much more practical hardware implementation and faster retrievals. In particular the project will explore four extensions of the Hamming memory: (i) the Grounded Hamming Memory, (ii) the Decoupled Hamming Memory, (ii) Two Level Decoupled Hamming Memory, and (iv) the Cellular Hamming Memory.