Please use this identifier to cite or link to this item: http://hdl.handle.net/2440/117927
Type: Thesis
Title: Long-Term Memory for Cognitive Architectures: A Hardware Approach Using Resistive Devices
Author: Wang, Peng
Issue Date: 2018
School/Discipline: School of Electrical and Electronic Engineering
Abstract: A cognitive agent capable of reliably performing complex tasks over a long time will acquire a large store of knowledge. To interact with changing circumstances, the agent will need to quickly search and retrieve knowledge relevant to its current context. Real time knowledge search and cognitive processing like this is a challenge for conventional computers, which are not optimised for such tasks. This thesis describes a new content-addressable memory, based on resistive devices, that can perform massively parallel knowledge search in the memory array. The fundamental circuit block that supports this capability is a memory cell that closely couples comparison logic with non-volatile storage. By using resistive devices instead of transistors in both the comparison circuit and storage elements, this cell improves area density by over an order of magnitude compared to state of the art CMOS implementations. The resulting memory does not need power to maintain stored information, and is therefore well suited to cognitive agents with large long-term memories. The memory incorporates activation circuits, which bias the knowledge retrieval process according to past memory access patterns. This is achieved by approximating the widely used base-level activation function using resistive devices to store, maintain and compare activation values. By distributing an instance of this circuit to every row in memory, the activation for all memory objects can be updated in parallel. A test using the word sense disambiguation task shows this circuit-based activation model only incurs a small loss in accuracy compared to exact base-level calculations. A variation of spreading activation can also be achieved in-memory. Memory objects are encoded with high-dimensional vectors that create association between correlated representations. By storing these high-dimensional vectors in the new content-addressable memory, activation can be spread to related objects during search operations. The new memory is scalable, power and area efficient, and performs operations in parallel that are infeasible in real-time for a sequential processor with a conventional memory hierarchy.
Advisor: Phillips, Braden
Liebelt, Michael
Ng, Brian
Dissertation Note: Thesis (Ph.D.) -- University of Adelaide, School of Electrical and Electronic Engineering, 2018
Keywords: Content-addressable memory
non-volatile memory
RRAM
ReRAM
cognitive architectures
long-term memory
semantic memory
hardware acceleration
in-memory computing
memristors
resistive devices
Provenance: This electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at: http://www.adelaide.edu.au/legals
Appears in Collections:Research Theses

Files in This Item:
File Description SizeFormat 
Wang2018_PhD.pdfThesis2.37 MBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.