Transcript Using Neural Networks in Decision Support Systems

Introduction
Core
Technology
Building and
Deploying
Neural
Networks
April 2005
Using Neural Networks in Decision Support Systems
Medical
Procedure
Certification
Crime
Forecasting
Jack Copper
NeuralWare
[email protected]
Hiroshi Maruyama
Grain Quality
Assessment
SET Software Co. Ltd.
[email protected]
© 2005 NeuralWare. All rights reserved.
P-1
NeuralWare
Since 1987, NeuralWare has created and marketed neural
network based Artificial Intelligence (AI) software for –



Data Mining (clustering)
Classification
Forecasting
NeuralWare collaborates with Customers and Partners to
Embed Intelligent Neural Network Engines into NextGeneration Products and Systems
© 2005 NeuralWare. All rights reserved.
P-2
Introduction
Characteristics of Neural Network Decision Support Systems

Integrate Data and Analytics

Adapt to Changing Conditions
© 2005 NeuralWare. All rights reserved.
P-3
Introduction
Benefits of Neural Network Decision Support Systems

Consistent Decisions

Rapid Decisions

Reproducible Decisions
© 2005 NeuralWare. All rights reserved.
P-4
Introduction
Examples of Neural Network Decision Support Systems

Medical Procedure Certification

Crime Forecasting

Grain Quality Assessments
© 2005 NeuralWare. All rights reserved.
P-5
Core Technology - Neural Networks
Output Layer
Target
Target
Decisions Based on Model Output
Model
Model
Hidden Layer
Input Layer
Historic Data
New Data
Artificial Neural Networks are connected hierarchies
of Artificial Neurons (also called Processing Elements)
© 2005 NeuralWare. All rights reserved.
P-6
Building Neural Networks
© 2005 NeuralWare. All rights reserved.
P-7
Evaluating Neural Network Performance
© 2005 NeuralWare. All rights reserved.
P-8
Evaluating Neural Network Performance
© 2005 NeuralWare. All rights reserved.
P-9
Deploying Neural Networks
Application Server Architecture
Server Contains Development and Run-Time Engine
Browser-based wired or
wireless remote PC clients do
not employ NeuralWare
technology
NeuralWare Technology (Run-Time
Engine/Models/FlashCode) embedded
in Server
© 2005 NeuralWare. All rights reserved.
P-10
Deploying Neural Networks
Distributed Intelligence Architecture
Server Contains Development Engine
Wired or wireless remote PC
clients employ embedded
NeuralWare technology (RunTime Engine/Models/FlashCode)
© 2005 NeuralWare. All rights reserved.
P-11
Case Study – Medical Procedure Certification
Objectives

Reduce Workload on Doctors and Registered Nurses

Improve Responsiveness to Customers (faster decisions)
Challenges

No “Gold Standard” for decisions – even Doctors sometimes disagree

Inconsistent data formats and labeling
Process

Used NeuralSight to build and evaluate ~ 30,000 Models in 3 weeks

Developed prototype software to permit altering Model decision threshold
© 2005 NeuralWare. All rights reserved.
P-12
Case Study – Medical Procedure Certification
Performance of best models
(ranked by Average
Classification Rate) for the
Global model and CT and MRI
Modality models
© 2005 NeuralWare. All rights reserved.
P-13
Case Study – Medical Procedure Certification
© 2005 NeuralWare. All rights reserved.
P-14
Case Study – Medical Procedure Certification
Acquire/Validate Case Input
Retrieve Metrics
Metric Database
Select/Execute Model
Apply Thresholds
Update Metrics
Process Manually
NO
Approve Procedure?
YES
YES
Selected for Audit?
© 2005 NeuralWare. All rights reserved.
DONE
P-15
Case Study – Crime Forecasting
Objectives

Identify Patterns in Criminal Activity that indicate Potential Future Trouble Spots

Redirect Police Resources to Focus on Areas where Serious Crime is expected to Increase
Challenges

Defining Crime Categories and Severity Levels

Inconsistent data formats and labeling; missing or non-existent data
Process

Used NeuralSight to Build and Evaluate ~ 10,000 Models in 1 week

On-going evaluation by researchers at Carnegie Mellon University
© 2005 NeuralWare. All rights reserved.
P-16
Case Study – Crime Forecasting
© 2005 NeuralWare. All rights reserved.
P-17
Case Study – Crime Forecasting
How to Forecast Change in Crime
Police know current crime levels
 Have allocated resources to respond to existing crimes
Most valuable information for tactical level planning:
 Where is crime likely to have large increases next month?
 Forecast crime by area and calculate:
Forecasted Change (t+1) = Forecast (t+1) – Actual (t)
The Benefit – Better Allocation of Scarce Resources
© 2005 NeuralWare. All rights reserved.
P-18
Case Study – Crime Forecasting
Forecasted Change for July
© 2005 NeuralWare. All rights reserved.
P-19
Case Study – Grain Quality Assessment
Objectives

Provide a Platform for rapidly and consistently assessing the quality of grain

Maintain detailed records of tests and build foundation for data mining
Challenges

No “Gold Standard” for decisions – even experienced human inspectors are inconsistent

Requires tedious work to identify wide variety of training data samples
Process

Used Predict and NeuralSight to Build and Evaluate many thousands of Models

Now developing image database to support agriculture research
© 2005 NeuralWare. All rights reserved.
P-20
Case Study – Grain Quality Assessment
An Instrument – and examples of seed images
© 2005 NeuralWare. All rights reserved.
P-21
Case Study – Grain Quality Assessment
Many (more than 300) initial features per seed
 Predict Variable Selection found a much smaller set of
features to use in building models
The characteristics of grain that are important are
difficult even for human inspectors to identify
 Multiple neural networks are used to make the hard decisions
The value of wheat and other commodities depends on
its quality – millions of dollars are at risk if quality
decisions are incorrect!
© 2005 NeuralWare. All rights reserved.
P-22
What Have you Learned?
Neural Networks make Powerful Decision Support Systems
 Human Judgment Determines the Cost/Benefit Tradeoff for Accuracy
Know your Problem !
Neural Network Decisions are based on Learning Patterns
 Relationships in Historical Data are the basis for Current Action
Know your Data !
© 2005 NeuralWare. All rights reserved.
P-23
Thank You !
Jack Copper
NeuralWare
[email protected]
Hiroshi Maruyama
SET Software Co. Ltd.
[email protected]
© 2005 NeuralWare. All rights reserved.
P-24