Thesis available: Neural model integrating episodic, semantic and temporal sequence memory

[Gerard J Rinkus]

FTP-host: cns-ftp.bu.edu
FTP-filename: /pub/rinkus/thesis_*.Z

The following Ph.D. Thesis is available via either anonymous ftp or my web
site (no hardcopies available). It is 249 pages long and the chapters can
be retrieved individually. (Specific retrieval instructions below).

===========================================================

Title:  A Combinatorial Neural Network Exhibiting Episodic and Semantic
          Memory Properties for Spatio-Temporal Patterns

                Gerard J. Rinkus
                Dept. of Cognitive and Neural Systems
                Boston University
                Boston, MA 02215
                rinkus at cns.bu.edu
                http://cns-web.bu.edu/pub/rinkus/www/

                                ABSTRACT

  This thesis describes TEMECOR (Temporal Episodic MEmory using
COmbinatorial Representations), an unsupervised, distributed, associative
network model of storage, retrieval and recognition of binary
spatio-temporal patterns, exhibiting episodic, semantic and complex
sequence memory. The original version of the model, TEMECOR-I, meets
several essential requirements of episodic memory-very high capacity,
single-trial learning, permanence (stability) of traces, and the ability to
store highly-overlapped spatio-temporal patterns, including complex state
sequences (CSSs) which are sequences in which the same state can recur
multiple times-e.g., [A B B A G C B A D]. Various parametric simulation
studies are reported, revealing that the model's capacity increases
faster-than-linearly in the size (i.e., number of nodes) of the network,
for both uncorrelated and correlated (specifically, complex sequence)
spatio-temporal pattern sets.

  However, TEMECOR-I fails to possess the crucial property that similar
inputs map to similar internal representations-i.e., continuity. Therefore
the model fails to exhibit similarity-based generalization and
categorization, which are the basis of many of those phenomena classed as
semantic memory.  A second version of the model, TEMECOR-II, adds the
property of continuity and therefore constitutes a single associative
neural network which exhibits both episodic and semantic memory properties,
and which does so for the spatio-temporal pattern domain. TEMECOR-II
achieves the continuity property by computing, on each time slice, t, the
degree of match, G(t), between its expected and actual inputs and then
adding an amount of noise, inversely proportional to G(t), into the process
of choosing a final internal representation at t.  This generally leads to
reactivation of old traces (i.e., greater pattern completion) in proportion
to the familiarity of inputs, and establishment of new traces (i.e.,
greater pattern separation) in proportion to the novelty of inputs.
Simulation results are given for TEMECOR-II, demonstrating the embedding of
similarity relationships in the model's adaptive mappings between inputs
and internal representations, and the model's ability to co-categorize
similar spatio-temporal events.

  The model is monolithic in that all three types of memory are explained
by a single local circuit architecture, instantiating a winner-take-all
network, that is proposed as an analog of the cortical minicolumn. Thus,
episodic (i.e., exemplar-specific) and semantic (i.e., general,
category-level) information coexist in the same physical substrate. A
principle/mechanism is described whereby the model's instantaneous level of
memory access-along the spectrum between highly specific (based on the
details of a single exemplar) and highly generic (based on the general
properties of a class of exemplars) memory access-can be controlled by
modulation of various threshold parameters.

===========================================================

FTP instructions:

(e.g. to retrieve chapter 1)

unix> ftp cns-ftp.bu.edu
Name: anonymous
Password: your full email address
ftp> cd pub/rinkus
ftp> get thesis_chap1.ps.Z
ftp> bye
unix> uncompress thesis_chap1.ps.Z

.then send to a postscript printer or previewer

Note: the file names and page lengths are:

thesis_chap1.ps.Z       ch. 1   49 (incl. prelim pages)
thesis_chap2.ps.Z       ch. 2   24
thesis_chap3.ps.Z       ch. 3   29
thesis_chap4.ps.Z       ch. 4  113
thesis_chap5.ps.Z       ch. 5    5
thesis_refs.ps.Z        refs.   11