Storing and retrieving data in a parallel distributed memory system
T. W. Potter · 1987
The storage and retrieval of patterns in a Hopfield parallel distributed memory is investigated experimentally with a view toward increasing its storage capacity. The first two Chapters give an overview of distributed memories and in particular the Hopfield distributed memory. This is followed by a Chapter which experimentally identifies the basic storage capacity of the original Hopfield memory when using text patterns. This dissertation then experimentally investigates new and untested methods to increase the storage capabilities of a Hopfield memory. Increasing the storage capacity by using the continuous-valued Hopfield memory is explored in Chapter 3 and the impact on capacity of data representation is experimentally investigated in Chapter 4. We then focus on new ways of storing data (changing the interconnect strengths) including in Chapter 7 developing a new method called Modifying the Energy Contour- MEC. In addition, this Chapter also outlines how to increase error-tolerance through the use of noisy patterns. The Hopfield memory is then contrasted to another intelligent memory subsystem based on more of a traditional computer technology. In Chapter 8 we see that traditional computer technology using data-parallel techniques has a greater storage efficiency than possible with current Hopfield distributed memories. The design of this data-parallel memory is based in part on what is learned experimentally from the preceding Chapters on the Hopfield memory. This fast data-parallel approach also supports retrieval of data patterns with noisy inputs although it does not have all the functionality of the Hopfield distributed memory. The following three results are the most significant outcomes of this dissertation. Experimentally, it was determined that: (1) The Hopfield memory during recall did a parallel, nearest-neighbor pattern search procedure. (2) The storage capacity of the Hopfield memory can be significantly improved but the storage efficiency is far less than data-parallel based associative memories. (3) A data-parallel implementation of the Nearest-Neighbor Rule provides for fast parallel search of pattern space and can support software-based learning procedures. This implementation can then behave as a Parallel Associative Memory dealing with inexact data in the recall key.