1 Abstract Natural Language Processing Confabulation Theory Texture Modeling Texture Classification Andrew Smith, Rupert Minnett, Soren Solari, Robert Hecht- Nielsen Confabulation Neuroscience Laboratory, University of California, San Diego, La Jolla, CA 92093-0407, USA Confabulation Theory [1,2] is a comprehensive theory of human and animal cognition, postulating that thinking is entirely symbolic processing. We provide a more in-depth look at Confabulation Theory, answering the following questions: What is cognitive knowledge? What is a thought process? How can symbolic processing be done by neurons? How can brains make probabilistic inferences? How can confabulation networks be simulated? [1] Hecht-Nielsen, R. Confabulation Theory. Springer-Verlag 2007. [2] Solari, S., Smith, A., Minnett, R., Hecht-Nielsen, R. “Confabulation Theory”, Physics of Life Reviews. (in press, 2008) Confabulating a plausible next sentence: After having been trained on a large corpus of consecutive sentence triples from newspapers, the Confabulation architecture is exposed to two sentences and generates a third. Input 1: Several other centenarians at Maria Manor had talked about trying to live until 2000, but only Wegner made it. Input 2: Her niece said that Wegner had always been a character – former glove model, buyer for Macy‟s, owner of Lydia‟s Smart Gifts downtown during the 1950s and „60s – and that she was determined to see 2000. Output: She was born in the Bronx Borough of New York City. This is a schematic of a Confabulation architecture for Natural Language Processing with a hierarchy of words, phrases, and sentences (from bottom to top). The words of two sequential novel context sentences are loaded into the red and brown modules and the network confabulates the green sentence. Confabulation is the universal basic operation of thought. • Confabulation is a simple, controlled winner- take-all competition between the symbols receiving excitation within a module. • Strongly active symbols are amplified and weakly active symbols are diminished by the thalamocortical attractor circuit. • This circuit is controlled by an externally supplied thought control signal. • The winning symbol maximizes “cogency” p( ), not a posteriori probability p( ) [1, 2]. • This affords a natural way to include contextual information in decision making. Knowledge links store all cognitive knowledge as pairs of meaningfully co- occurring symbols. • An active (firing) symbol in one module delivers excitation to a symbol in another module through this unidirectional link. • Knowledge links are learned through repeated exposure to co-active pairs of symbols (Hebbian learning). • The set of knowledge links from one module to another is termed a knowledge base. • The strength of the knowledge link from symbol a to symbol b is logarithmically related to the conditional probability p(a | b). A Thalamocortical module and its symbols describe exactly one attribute of a mental object (e.g. a visual or auditory object, language unit, movement or thought process, plan). • Symbols are encoded as sparse populations of neurons within a module. • Each population is a stable state of a neuronal attractor network (the cortex-thalamus circuit). • The set of symbols within a module enumerate the possible descriptors of that module‟s attribute (e.g. colors, tactile textures, scents, words). • The human cerebral cortex is divided into thousands of localized patches, which, with its paired thalamic region, constitutes a module. Confabulation Theory – The Mechanism of Thought Above are six novel textures, two from each of the three classes. Humans can easily identify the similarities and correctly classify the textures. To the right are the results of a Confabulation architecture trained on the associations between textures and their classes and presented with a novel stimulus. In this example, four source modules each have one active symbol ( ), and provide input to the symbols of a target module. The combined excitation defines a distribution on the target symbols. The thought control signal initiates the confabulation operation, which selects symbol 9 because its external excitation is strongest. Symbols are sparse populations of neurons. Neurons may be shared between symbols, though significant overlap is vanishingly unlikely. Fabric Stone Wood F: 18.60% S: 66.94% W: 14.46% F: 4.84% S: 7.96% W: 87.20% F: 21.67% S: 60.98% W: 17.35% F: 7.35% S: 6.37% W: 86.28% F: 90.27% S: 8.30% W: 1.43% Robust texture classification is largely an unsolved problem that can benefit directly from biologically inspired approaches such as Confabulation Theory. • A multi-resolution Gabor filter bank is applied to the image, extracting features similar to those in the visual cortex. • Symbols are created from sparse vectors of the strongest filter responses within small training texture patches. • Knowledge links are learned from the texture processing module to the class label module. • Novel stimuli are presented and the system attempts to infer the correct class label. Stimulus: Classification: Performance: Here, links between frequently co-occurring symbols have been learned. When viewing an apple, the strongly linked symbols in other modules are excited, and a complete mental representation of the apple is evoked. Input 1: Michelle strengthened from a Category 2 to a Category 4 storm Saturday, with winds reaching 140 mph, but it was expected to weaken before it reached Florida. Input 2: The storm or its effects could strike the Keys and South Florida tonight or early Monday, said Krissy Williams, a meteorologist at the National Hurricane Center in Miami. Output: Forecasters warned residents to evacuate their homes as a precaution. Input 1: He started his goodbyes with a morning audience with Queen Elizabeth II at Buckingham Palace, sharing coffee, tea, cookies and his desire for a golf rematch with her son, Prince Andrew. Input 2: The visit came after Clinton made the rounds through Ireland and Northern Ireland to offer support for the flagging peace process there. Output: The two leaders also discussed bilateral cooperation in various fields. Learning Generation Application: texture reconstruction Future work with visual confabulation: • higher levels of abstraction using a symbol hierarchy • image super-resolution / image enhancement • object detection, classification, and recognition • learning links between images and words for: • automatic image annotation • content-based image retrieval Apply filters to training set (Gabor filters with 8 orientations, 5 scales). Collect frequently occurring filter response vectors to form a symbol lexicon for each scale. The Gabor representation is formed by simple look- up and substitution, since the symbol lexicon constitutes a codebook of filter responses. Generate an image formed from filter coefficients using any regression technique (e.g. neural network, least-squares). Learn knowledge links between neighboring symbols. Given context (e.g. neighbors), confabulate plausible symbols, choosing a mutually self-consistent set (maximally cogent). learned symbols Gabor filters training data Given an image with a hole in its symbol representation, can we confabulate a plausible replacement? ? ? ? ? … … scale 2 scale 3 scale 4 Each of the empty modules confabulates a maximally cogent symbol, based on knowledge links from its neighbors. original damaged reconstructed some real-world textures This experiment uses 596 sets of knowledge links:
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1
Abstract
Natural Language Processing
Confabulation Theory Texture Modeling
Texture Classification
Andrew Smith, Rupert Minnett, Soren Solari, Robert Hecht- NielsenConfabulation Neuroscience Laboratory, University of California, San Diego, La Jolla, CA 92093-0407, USA
Confabulation Theory [1,2] is a comprehensive theory of human and animal
cognition, postulating that thinking is entirely symbolic processing. We provide a more
in-depth look at Confabulation Theory, answering the following questions:
What is cognitive knowledge?
What is a thought process?
How can symbolic processing be done by neurons?
How can brains make probabilistic inferences?
How can confabulation networks be simulated?
[1] Hecht-Nielsen, R. Confabulation Theory. Springer-Verlag 2007.
[2] Solari, S., Smith, A., Minnett, R., Hecht-Nielsen, R. “Confabulation
Theory”, Physics of Life Reviews. (in press, 2008)
Confabulating a plausible next sentence: After having been trained on a
large corpus of consecutive sentence triples from newspapers, the Confabulation
architecture is exposed to two sentences and generates a third.
Input 1: Several other centenarians at Maria Manor had talked about trying to live until
2000, but only Wegner made it.
Input 2: Her niece said that Wegner had always been a character – former glove model,
buyer for Macy‟s, owner of Lydia‟s Smart Gifts downtown during the 1950s and
„60s – and that she was determined to see 2000.
Output: She was born in the Bronx Borough of New York City.
This is a schematic of a Confabulation
architecture for Natural Language
Processing with a hierarchy of words,
phrases, and sentences (from bottom to
top). The words of two sequential novel
context sentences are loaded into the red
and brown modules and the network
confabulates the green sentence.
Confabulation is the universal basic
operation of thought.
• Confabulation is a simple, controlled winner-
take-all competition between the symbols
receiving excitation within a module.
• Strongly active symbols are amplified and
weakly active symbols are diminished by
the thalamocortical attractor circuit.
• This circuit is controlled by an externally
supplied thought control signal.
• The winning symbol maximizes “cogency”
p( ), not a posteriori probability
p( ) [1, 2].
• This affords a natural way to include
contextual information in decision making.
Knowledge links store all cognitive
knowledge as pairs of meaningfully co-
occurring symbols.
• An active (firing) symbol in one module delivers
excitation to a symbol in another module through
this unidirectional link.
• Knowledge links are learned through repeated
exposure to co-active pairs of symbols (Hebbian
learning).
• The set of knowledge links from one module to
another is termed a knowledge base.
• The strength of the knowledge link from symbol a to
symbol b is logarithmically related to the conditional
probability p(a | b).
A Thalamocortical module and its symbols describe exactly one attribute of a mental object (e.g.
a visual or auditory object, language unit, movement
or thought process, plan).
• Symbols are encoded as sparse populations of neurons
within a module.
• Each population is a stable state of a neuronal attractor
network (the cortex-thalamus circuit).
• The set of symbols within a module enumerate the possible
descriptors of that module‟s attribute (e.g. colors, tactile
textures, scents, words).
• The human cerebral cortex is divided into thousands of
localized patches, which, with its paired thalamic region,
constitutes a module.
Confabulation Theory –The Mechanism of Thought
Above are six novel textures, two from
each of the three classes. Humans can
easily identify the similarities and
correctly classify the textures. To the right
are the results of a Confabulation
architecture trained on the associations
between textures and their classes and
presented with a novel stimulus.
In this example, four source modules each have one active
symbol ( ), and provide input to the symbols of a target
module. The combined excitation defines a distribution on the
target symbols. The thought control signal initiates the
confabulation operation, which selects symbol 9 because its
external excitation is strongest.
Symbols are sparse populations of neurons.
Neurons may be shared between symbols,
though significant overlap is vanishingly unlikely.
Fabric
Stone
Wood
F: 18.60%
S: 66.94%
W: 14.46%
F: 4.84%
S: 7.96%
W: 87.20%
F: 21.67%
S: 60.98%
W: 17.35%
F: 7.35%
S: 6.37%
W: 86.28%
F: 90.27%
S: 8.30%
W: 1.43%
Robust texture classification is largely an
unsolved problem that can benefit directly from
biologically inspired approaches such as
Confabulation Theory.• A multi-resolution Gabor filter bank is applied to the
image, extracting features similar to those in the visual
cortex.
• Symbols are created from sparse vectors of the strongest
filter responses within small training texture patches.
• Knowledge links are learned from the texture processing
module to the class label module.
• Novel stimuli are presented and the system attempts to
infer the correct class label.
Stimulus: Classification: Performance:
Here, links between frequently co-occurring symbols have
been learned. When viewing an apple, the strongly linked
symbols in other modules are excited, and a complete
mental representation of the apple is evoked.
Input 1: Michelle strengthened from a Category 2 to a Category 4 storm Saturday, with
winds reaching 140 mph, but it was expected to weaken before it reached
Florida.
Input 2: The storm or its effects could strike the Keys and South Florida tonight or early
Monday, said Krissy Williams, a meteorologist at the National Hurricane Center
in Miami.
Output: Forecasters warned residents to evacuate their homes as a precaution.
Input 1: He started his goodbyes with a morning audience with Queen Elizabeth II at
Buckingham Palace, sharing coffee, tea, cookies and his desire for a golf
rematch with her son, Prince Andrew.
Input 2: The visit came after Clinton made the rounds through Ireland and Northern
Ireland to offer support for the flagging peace process there.
Output: The two leaders also discussed bilateral cooperation in various fields.
Learning Generation Application: texture reconstruction
Future work with visual confabulation:
• higher levels of abstraction using a symbol hierarchy
• image super-resolution / image enhancement
• object detection, classification, and recognition
• learning links between images and words for:
• automatic image annotation
• content-based image retrieval
Apply filters to training
set (Gabor filters with
8 orientations,
5 scales).
Collect frequently
occurring filter
response vectors
to form a symbol
lexicon for each
scale.
The Gabor representation
is formed by simple look-
up and substitution, since
the symbol lexicon
constitutes a codebook of
filter responses.
Generate an image formed
from filter coefficients using
any regression technique (e.g.
neural network, least-squares).
Learn knowledge
links between
neighboring
symbols.
Given context (e.g.
neighbors), confabulate
plausible symbols, choosing
a mutually self-consistent
set (maximally cogent).
learned symbols
Gabor filters
training data
Given an image with a hole in
its symbol representation, can
we confabulate a plausible
replacement?
?
?
?
?
…
…
scale 2
scale 3
scale 4
Each of the empty
modules
confabulates a
maximally cogent
symbol, based on
knowledge links
from its neighbors.
original damaged reconstructed
some real-world textures
This experiment
uses 596 sets of
knowledge links:
Related Documents
