The MemeStore concept is based on the way humans and
animals store memories.
MemeStore is an adaptive knowledge base, which
may be given a personality, through
the use of
personality traits.
The whole concept is based on animals who evolved, with the
imperative to survive and
reproduce. These imperatives are controlled through hormones
and
memories, creating
attractions and aversions --
related to danger and safety, danger being
anything to do with
predation or any other
type of harm; and safety being food, shelter, play, and
reproduction --
which are stored as memories, and then
drive
certain types of behaviour; or
predispose
the
animal to certain types of behaviour -- see
Memestore -
a Knowledge Base.
Memestore running on a
30/40 Core Computer with 64G of Memory
18/28 graph traversal bots with one per core.
1 million word/node knowledge base.
500/1000 bytes/node => 1
Gigabyte - Maximum 10G.
1 Gigabyte for visualization data - Maximum 10G.
1 Gigabyte as working scratch memory - Maximum 10G.
12 Memestore Modules:
graph
traversal,
character
recognition,
scanning/reading,
search-engine,
semantic
command mapper,
memory trails,
memory stores,
query/command,
personality
traits/attributes,
reward system,
conversation,
visualization
Module 1 - Graph Toolkit
Graph library to build the knowledge base.
Granular locks are used to lock sections of the graph.
Node/Vertex:
word(char,64),
concept(uint),
locale(uint),
type(uint/isa,hasa,etc.),
weight(uint),
like(uint), dislike(uint),
vertex(uint,200/1000), search-vertex(uint)
Module 2 - OCR and Vision
A way to read characters using a character recognition
engine (OCR),
The vision part of this module, consist of a hardware and
software system,
which is able to identify living and non-living entities;
as well as, perceive
and understand a living entity's behaviour and its
reactions to stimuli.
Module 3 - Scanning/Reading
Updated content engine with dictionary reading rules.
Finite state machine generator using XML-based rule sets.
Read from encrypted databases and rule sets.
Use grammar, thesauri, and dictionaries.
Module 4 - Search Engine
Mapping of memory trails to user queries and scanned
pages.
Dynamic updates of relevancy of scanned pages using
signatures of memory trails.
Use XML-based rule sets. The
search module is based on our CETE engine.
Use grammar, thesauri, and dictionaries.
Module 5 - Semantic Command
Mapper
Mapping of natural language commands to pseudo or
hardware-based commands,
Use semantic command map with actor, patient, subject,
etc..
Module 6 - Memory Stores
Access, copying, retrieval, update for both temporary and
permanent stores.
Rules for increasing and decreasing
weight/relevance/like/dislike of stored memory.
Data driven module with lexer and parser.
Module 7 - Memory Trails
A logging mechanism to store traversals of the knowledge
base graph.
They are stored as signatures and can be overlayed on top
of the graph,
without changing the graph, They are referred to as path-overlays.
Capability to retrieve a specific trail and its
associations, see white-paper.
Module 8 - Query/Command
Natural language query module for accessing the memory stores.
This module relies on the OCR/vision
and the reading/scanning module.
This module and the semantic
command mapper may be joined.
Test by generating a natural language
script, with t-script.
Module 9 - Personality
Traits
A database of traits and their associated concepts, stored
as a graph.
Link set of commands to execute, seen as observable
behaviour, that a person could interpret.
Link concepts of character traits, aversions, attractions
with stored memories.
Module 10 -
Behaviour/Reward Engine
Increase/decrease of the relevance of stored memories, is
based on behaviour,
which is observed or effected by the MemeStore. Several criteria
are to be used,
which are: frequency, user reinforcement, etc..
The engine is built using propositional calculus; and
modeled using graph theory, and
De Brujin sequences. A level of fuzzy is added, consisting of
the two meters dealing with like
and dislike. The
propositional
calculus
engine, of the
MemeStore, can
ultimately
be
burned onto an integrated
circuit.
Our MemeStore
PCG engine cannot be
circumvented, by building a similar propositional
calculus engine, using Bayes theorems, formal
grammar or graph theory methodologies.
Module 11 - Conversation
Conversing with the MemeStore relies on voice recognition and vision
through
the use of a vision system, a speech recognition engine,
and a speech engine.
The conversation
module is built on top of the query/command module.
The OCR/vision
engine, as well as, the speech recognition engine
rely on
the scanning/reading
or query/command
module, as back-end to their function.
Both engines output text, that is fed into the scanning/reading
module; which then
gauges, whether or not, the inputted text is meaningful;
that is, relates to concepts
and notions that are understandable by the MemeStore.
The conversation module is modeled on the way most people
carry a conversation.
They are constantly accessing their knowledge base, and
weighing the relevance, of
the information, based on their personality profile. A
simple example would be carrying
a conversation in the context of a reception. One would
choose to start a conversation
with a person whose attire is attractive; and, then one
would choose to continue or
terminate the conversation, depending on the relevance and
attractiveness of what is
said by the other party. The manner in which a
conversation is terminated again depends
on a person personality profile and the context of the
situation -- experiential memories.
Module 12 -
Visualization
A way to visualize the MemeStore in action, by overlaying the memory
trails in
a heat color-coded fashion, on top of a 3D sphere-like
image. For search-engines,
the data is overlayed on top of a 3D view of the world, if
the engine is world accessible.
Module 13 to N
Modules can be added to the MemeStore to extend it, or
build a humanoid robot.
Such modules are: a manipulation module, a data
mining and analysis module,
a gaming
module, a generic problem solving module, a learning
by mimicry
module, etc..
The CETE Engine
After building a semantic network with natural language
relationships between keywords;
the CETE search engine
allows the following to be done:
Indexing
- Index file-name of documents for
path specification (path-spec) queries.
- Index keywords found in pages and
documents for keyword queries.
- Content analyze the unstructured
text in the pages and documents, using our statistical
natural language processing (NLP) approach.
- Build signatures of every set of
extracted keywords and their relationships. These
signatures are called semantic signatures.
- Build signatures of the path
traversed, by every set of extracted keywords, in the
semantic network. These signatures are called network
path signatures.
- Associate semantic and network
path signatures with scanned pages and documents.
- Sort and log the signatures for
retrieval of scanned pages and documents.
Search Queries
- Make searches using
path-spec (file-name keywords) queries.
- Make searches using keywords only
(i.e. clustering legacy way).
- Make searches by traversing the
semantic network looking for relationships between
query keywords.
- Build a graph of the relationships
between query keywords (semantic signatures).
- Extract network path signatures
from these semantic network traversals.
- Sort and log network paths and
semantic signatures for retrieval of scanned
pages and documents.
- Track user behaviour (desktop
clicking, voice, eye movement, etc.) to modify the
strength of these network paths.
- Return results by comparing the
query signatures with the stored network-path and
semantic signatures.
- Return results with just the
strengthened network paths which refer to files that
were deemed to satisfy users.
- Build semantic network path
overlays, using the extracted paths for visual
feedback; for example, a different color (e.g.
heat-coded), depending on how satisfied, users were
with the results of the query; or how strong the
relationships between keywords, and the concepts to
which they relate, are.
- Build networks based on path-spec
keywords that can be displayed to users interested, in
what the collection of keywords they use, in
specifying their file names, look like in a graph.
Patent
Pending
Tsert.Com