Transcript In Microsoft Power Point Format

Incremental Learning
With Neural Networks
February 1, 2001
May01-14 Team Members
Clients/Faculty Advisors
David Herrick
Dr. Eric Bartlett
Brian Kerhin
Chris Kirk
Ayush Sharma
Overview
• Problem Statement (Ayush)
• System Overview (Dave)
• Design Objectives (Ayush)
• Technical Design (Dave)
• Intended Users (Ayush)
• Evaluation of Project Success (Chris)
• End-Product Description (Brian) • Possible Future Work (Chris)
• Assumptions/Limitations (Brian) • Human/Financial Budget (Chris)
• Project Risks & Concerns (Brian) • Lessons Learned (Chris)
• Technical Approach (Brian)
• Closing Summary (Chris)
ANN Description
=
Real neurons are decision making cells in the brain
Software neurons function similarly, to make decisions in software
The brain is a network of “Real neurons” making decisions in parallel
ANNs accomplish the same goal using “Software neurons”
ANNs are used to interpolate nonlinear systems that are very complex
Predict trends in the Stock Market
Predict trends in power use through out the year
Predict trends in global whether patterns

Problem Statement
Unlike humans, who incrementally learn information as it is introduced,
ANNs learn all at once. An existing ANN cannot adapt to dynamically
changing system (i.e. CATASTROPHIC FORGETTING). Hence, as a
system changes, a new ANN must be created from scratch. For many reallife applications of ANNs, it is impractical to regularly create replacement
ANNs.
Design Objectives
• Create a software design document
• Create an incrementally learning ANN
• Create a GUI for incrementally learning ANN
• Apply to a power load problem and compare to traditional ANNs
Hardware and Software (Supplies)
• Hardware
• Intel Gateway PCs in Adaptive Computing Laboratory
• Software
• Microsoft Windows NT4.0
• Microsoft Visual C++
• Microsoft Visual Basic
Intended Users
• Adaptive Computing Laboratory
• Dr Eric Bartlett
• Research Assistants
• Other Neural Network Programmers
End-Product Description
• Incrementally learning ANN to model dynamic nonlinear systems
• Learning System can learn new data without forgetting
• Has goodness measures
• GUI operated, via command line
Assumptions
• User should have a basic knowledge of neural networks
• Data file will be tab delimited following Dr Bartlett’s
specified file format
Limitations
•
Set time period to produce code and documents
•
Number of developers is not based on commercial properties
of the software
•
Computationally intensive technique of problem solving
Project Risks & Concerns
• Major accidents can (and have) cripple team members
• Finishing on time
• Incremental learning may not improve goodness measures or
error results
Technical Approach
(Design Alternatives)
• One strong neural network with augmenting neural nets
• Several augmenting neural nets
• Update a single augmenting neural net
•Produce result based on outputs average
•Produce result based on output sums
• Multiple weak neural nets average results
Incremental Learning System
System Overview
Technical Design
Technical Design
Incremental Learning System
System Overview
Evaluation of Project Success
1st Semester Milestones
• Project Plan (Fully Met)
• Project Poster (Fully Met)
• Design Report (Fully Met)
• Hardware Requirements (Fully Met)
• Software Requirements (Fully Met)
• Software Design (Fully Met)
Evaluation of Project Success (cont.)
2nd Semester Milestones
• Software Implementation (Partially Met)
• 100% of algorithm designed
• 10% of algorithm implemented
• Final Implementation (Not Met)
• Final Report (Not Met)
• Presentation for Industrial Review Panel (Partially Met)
• 80% of Presentation Completed
Recommendations for Further Work
• Explore alternative learning styles
• More research tools in the field of artificial intelligence
Human Budget
Personnel
Estimated Total
(1st + 2nd Semesters)
Actual (to date)
Ayush
125 hours
86 hours
Brian
115 hours
73 hours
Chris
140 hours
91 hours
Dave
130 hours
88 hours
TOTAL
510 hours
334 hours
Financial Budget
Items
Estimated
Actual
Poster
$50
$35
Hardware
$0
$0
Software
$0
$0
Books
$100
$0
Total
$150
$35
Lessons Learned
• Establish two weekly meetings
• Progress meeting with faculty advisors
• Development meeting with team members
• Wounded team members don’t improve group efficiency
• Keep faculty advisors well-informed of progress and seek feedback
• Plan to complete milestones ahead of schedule
• Balance workload among team members
Closing Summary
• New way of looking at neural networks
• Overcomes limitations of traditional neural nets
• Can be greatly reused and built upon
• Will further the field of Artificial Intelligence
Questions?