Transcript Slide 1
MODELING AND VISUALIZING
DYNAMIC ASSOCIATIVE
NETWORKS:
Towards Developing a
More Robust and
Biologically-Plausible
Cognitive Model
Based on Dr. Anthony Beavers’
ongoing research
By Michael Zlatkovsky, dual-major in
Computer Science and Cognitive Science
I’M A PC...
I’m a neural net
WHY NEURAL NETS?
Pattern recognition
Inferring a function by observation
Robustness against errors
Parallel nature
ARTIFICIAL NEURAL NETWORKS
ARTIFICIAL NEURAL NETWORKS
Artificial way of adjusting: setting weights
DR. BEAVER’S DYNAMIC ASSOCIATIVE
NETWORK MODEL
Dr. Beavers,
Director of UE’s
Cognitive
Science
Department, is
attempting to
explore a
different model
of cognition.
DR. BEAVER’S DYNAMIC ASSOCIATIVE
NETWORK MODEL
No more mystery
“hidden layer”
Learning through the
order and structure of
experience
No “unnatural”
training
Organic network
Can incorporate new
information
DAN’S COGNITIVE ABILITIES COME FROM
LONG-TERM LEARNING AND CURRENT STATE
TRANSLATION INTO A NODE-CENTRIC
MODEL
EARLY EXCEL PROTOTYPE
THE DAN SOFTWARE SUITE
Based on prototype, create a self-contained DAN
Model
Written in Java;
object-oriented approach
Expand on features of Excel Model
(various activation modes, learning
mode, settings)
Most importantly: focus on
design fundamentals to ensure
speedy operation and high
capacity.
Create visualization routines
RE-CALCULATIONS
Most frequent operations
DANs are massively parallel
Re-computing from scratch: O(n2).
EX: for 1000 node-network, change
in 2 nodes that impact 5 nodes each...
Instead of 10 re-calculations, 1,000,000!
My scheme: buffered change-propagating
dependency-driven re-calculations
OTHER DESIGN CONSIDERATIONS
General separation of concerns (59 classes)
Model-View-Controller
“Core framework” with “helper” controllers &
GUI views/wrappers
GUI look, cross-platform
VISUALIZATION
PREFUSE
framework
Radial tree layout
(PREFUSE)
Color nodes based on
activation
Color edges based on
connection type
Highlighting,
animation, etc.
RESULTS: DAN SOFTWARE SUITE
Overall successful
Quick
Convenient UI
Adaptable
True to model
RESULTS: DAN MODEL
Promising results: various rudimentary
cognitive abilities:
“Initial Intelligence”: pattern recognition,
feature detection, memorization of simple
sequences, identification of similarities and
differences, storage of relational data,
comparison and classification, etc.
Possibly, building blocks of more sophisticated
intelligence.
RESULTS: DAN MODEL
Has not gone unchanged:
RESULTS: DAN MODEL
Has not gone unchanged:
RESULTS: DAN MODEL
Has not gone unchanged:
training:
“the boy woke up”
“the boy fell asleep”
“the boy woke up”
“the boy fell asleep”
RESULTS: DAN MODEL
Has not gone unchanged:
training:
“the boy woke up”
“the boy fell asleep”
“the boy woke up”
“the boy fell asleep”
RESULTS: OVERALL
More robust?
Don’t know... Yet.
Received with curiosity and some enthusiasm
by researchers working in the field.
More biologically plausible?
Absolutely.
Hebbian Neurological Principle: nodes that
“fire together, wire together”.
Contrast with ANNs’s statistically-based
learning
I’m a DAN