The Conceptor as a Model for Speech and Consciousness Emergence

1
The Conceptor as a Model for Speech and
Consciousness Emergence
Deriving Precise Concepts from Imprecise Sensing
Boleslaw K. Szymanski and Konrad
Fialkowski Rensselaer Polytechnic Institute Troy,
NY 12180, USA email fialkows_at_aol.com,
szymansk_at_cs.rpi.edu June 3, 2003 3. Kongres
Informatyki
2
The Conceptor A Network of Interconnected
Concept Nests

• Purpose automatic knowledge creation and
  extraction from sensor input
• Means an attribute-based machine that
  automatically generates concepts
• from sensor collected data
  and/or other concepts
• Processing deriving a coherent and condensed
  representation of its
• universe (environment)
  exclusively from sensor measurements
• of the universe in a solely
  recursive manner.
• Concepts dynamic objects evolving from concept
  seeds, that are initiated
• at any level of conceptor
  processing based on pattern
• repeated appearance
• Concept interconnections weighted links
  established and continuously
• adjusted between related
  concepts.
• Both the concepts and the connections are
  dynamically adjusted by new information received
  from its universe (environment).

3
Attribute Transformers

• Attributes classes of environment properties
  (e.g. shape, color)
• measurable by a sensor
• Attribute values discrete descriptors of the
  attribute in the part of a scene
• observed and coded by a sensor
  (e.g., triangle, circle or blue)
• Scene a set of attribute values produced by all
  sensors at the given time
• instant.
• The granularity of sensor coded values, the
  frequency with which sensors collect values and
  of attributes recognized by the sensors define
  the universe of the conceptor.
• Extended Rough Set Cells units that process
  subsets of scenes identifying repeating patterns
  of corresponding attribute values (such patterns
  becomes seeds of concepts of the lowest level).

4
Sample Set of Scenes 15 scenes each with 7
descriptors
Time sequence enhanced scene 2-4 o b g c b m g i
a
5
Concepts of the higher order(s)
Each identified concept seed is given a unique
coded name and its own concept nest is
established. In the example, three concepts would
be established and named ? for the concept
lt"b/a2","a/a4","s/a5","e/a7"gt ? for the concept
lt"i/a1","m/a3","a/a4","g/a6","e/a7"gt ? for the
concept lt"r/a2","o/a3","s/a4","t/a5","e/a6","r/a7"
gt.
Set of scenes after coding
The input scene is then replaced by the input in
which vectors containing identified concept seeds
are coded with their names and processing
continues in search of higher level concepts.
6
Functions of the Concept Nest

• Refine and Update each nest uses Mask Matrix
  (MM) to filter all
• confirming or
  nearly confirming scenes in each new
• scene (fuzzy
  filtering). MMs are bus connected.
• Split a Nest when two concepts are detected in
  a nest, a new
• accompanying
  concept nest is created.
• Consolidate Two Nests when two nests are
  discovered to represent the
• same concept,
  they are merged into one.
• Link evaluate and adjust concepts links with
  other concepts based on the
• content of
  each nest (stored scenes) or class structure.
• The two functions splitting and consolidating,
  are important for dynamically changing knowledge
  (initial learning, reeducating, functioning in
  new environments).
• An interesting feature of the conceptor is that
  it cannot unlearn once acquired knowledge, it can
  only reevaluate it. Hence, it is better to let
  old conceptor die while educating a new one, when
  the environment changes.

7
Conceptors Directives and Implementation

• Illumination activating of all concepts related
  to the concepts in a Boolean
• expression parameter of the
  illumination directive. It reveals
• both the character and
  strength of the relations.
• Gedanken Experiment a conditional directive
  with if-clause that answers
• queries like what would the
  representation be, if two unrelated
• concepts were related
  (e.g., they were equivalent).
• Detection of Cause-Effect Relations processing
  of scene attributes with
• recording time correlation
  between descriptors of different
• scenes.
• Implementation uses networked processors and may
  use parasitic computing by broadcasting the
  scenes and parity check of the packets filtering
  them to the relevant concept nests.
• The scalability of the approach can be derived
  inductively from the uniformity of the conceptor
  design at different processing levels and the
  successful simulation of the small size conceptor.

8
The Conceptor An Architecture and Operations
Environment
Sensors
Sensor Measurement Coding
Scene Collection and Discovery
Bus
Concept Filtering
Command Source (Illuminations)
Concept Processing
9
An Illumination as a Model of Consciousness
Basic Properties

• Properties of Consciousness postulated in the
  literature
• Mode of action changing state right where it is
  by acquiring some property
• (Dennet, 1992) illumination is
  created in response to the input
• command
• Searchlight approach (Crick, 1984)
  illumination is performed by activation
• of a group of nests and is the result
  of processing and not processing
• itself (Jackendoff, Johnson-Lairds)
• Isotropy anything learned can contribute to
  anything that is currently
• confronted (Fodor, 1983), growing
  concepts clearly posses this
• ability
• Distinction between seeing and identifying
  (Treisman, 1988) where
• seeing is done by sensors and
  identifying is accomplished by the
• the concept nests
• Unconscious execution of simple or over-learned
  tasks (Dennet, 1992)
• can be achieved for the scenes that
  perfectly match a concept nest

10
Interesting Analogies with Consciousness

• The role of sleep processing new scenes in
  real-time precludes executing
• nest functions
  which are essential for the creation of new
• knowledge and
  efficient processing of data. Hence,
• dormant input
  state (sleep) is used for the latter.
• Speech recognition recent research (Science,
  2000) indicates that
• newborn humans
  and monkeys respond to speech
• similarly, yet
  language skills develop only in humans
• Language recognition associating single word
  with a concept is possible
• for animals, yet
  higher level concepts over a set of
• concepts and
  links between concepts ( in the form
• concept A
  relation R concept B ) is a basis for
• illumination
  (conceptor consciousness) as well for
• language
  recognition, which is likely to emerge
• during an early
  childhood development.
• .

11
Open ProblemsCurrent Research Issues

• Sensitivity to attributes granularity.
• Merging a range of values into a
  mega-value of an attribute increases
• chances of merging concepts, but
  looses precision of a world-view.
• Minimum size requirement for non-trivial
  results.
• We estimate that for a meaningful
  processing, tens of thousand of
• concept nests needs to be created,
  each with its own processor and
• memory, hence, non-trivial computer
  resources are needed
• Aging of the conceptor.
• The conceptor never forgets seen
  scenes, hence when its world
• changes drastically, it becomes less
  useful, when to replace it?
• Pre-wiring conceptors learning and sharing its
  knowledge.
• By selecting the level of
  sophistications of attributes, some knowledge
• can be inherited by the conceptor and
  invoking illuminations, forces
• the conceptor to share knowledge to
  some extent. More is needed.
• Logical reasoning by the conceptor.
• How to pre-wire first or higher order
  logic into the conceptor?

12
Conclusions Unique Characteristics of the
Conceptor

• Three levels of processing.
• Interfacing the environment and processing and
  coding attributes
• Pattern discovery is done inductively by ERSCs
  and pattern extraction is accomplished in
  parallel by each nest MM
• Concept processing, splitting, merging and
  illumination, is done by
• concept nests (but illumination is
  invoke at CC)
• Fully automatic, unsupervised extraction of
  abstractions from input
• Concept nests are created solely by
  repetitive sets of attribute values
• Formulation of precise concepts from imprecise
  input
• Each concept is a set of scenes in
  which exact or slightly distorted
• values of the concept attributes
  appear
• Abstracting across spatial and temporal
  dimensions
• Descriptor co-appearance applies to a
  scene a sequence of scenes
• Using illumination for hypothesis and analogy
  creation
• Gedanken experiment tests hypotheses
  and activating similarly related
• concepts in certain kinds of
  illuminations enables analogy discovery