Defining Constituents, Data Vizzes and Telling a Data Story
Watson Guiding The Way Using cognitive computing to discover the secrets behind successful guide dogs
1. Watson Guiding The Way
Using cognitive computing
to discover the secrets behind
successful guide dogs
Jane Russenberger, BS
Guiding Eyes for the Blind
Yorktown Heights, NY
October 24, 2016
2. • Connect people with vision
loss with exceptional dogs to
enjoy greater independence
• Dogs placed at no charge,
worldwide
• 60 years of service
• >7,300 teams graduated
• 501 (c ) (3) non-profit
• Yorktown Heights, NY
4. Goal 12% more dogs become guides
49%
success rate
today
$920K invested
means success
Release
Success
5. Lots of data on
>1200 dogs
Puppy
raising
In for
training
Puppy @
8 weeks
Training
“By collecting information about our dogs,
we can dig into the data to pull out meaningful insights
about health, behavior, temperament and more.”
—Thomas Panek, President and Chief Executive Officer,
Guiding Eyes for the Blind
6. 6
Phenotype = Genotype + environment
P= G + E
Interaction between genetics and environmental factors
P= f(G,E)
Environment
80%
Genetics
20%
7. Genetic Differences
Poor scores on pup test
Failed guide dog training
Good scores on pup test
Successful as a guide dog
Axel Anders
8. • Not adjusting approach
based on dog’s arousal level
• Confusing communication
Experienced raiser:
resolves stress
• Adjusts approach
• Clear communication
New raiser:
unresolved stress
9. Dr. Chris Tseng
Professor and
Director of
Data Intelligence Lab
for Cognitive
Computer Learning
Jane Russenberger, BS
Senior Director of
Genetics and Breeding
10. Comments
on Training &
Socialization
Puppy
Raiser
Reports and
Narrative
Success or
Failure of
Dog
• 100% accuracy
• All 105 record used
Success of
Dog
• 100% accuracy
• 67/105 records
used
Success of
Volunteer
Puppy Raiser
Results- New Data
Supervised Learning
Project 1- Prediction
11. Results
Increased success Reduced costs
expected
New insights into
the personalities
means happy volunteer puppy
raisers, more people matched
with guide dogs to help meet
growing demand
through greater efficiency in
the guide dog program, from
breeding to placement with a
human in need
of dogs and matching to the
volunteer puppy raisers
13. Louis Freund PhD
Professor and Director,
Graduate Program in
Human Factors and
Ergonomics
Department of
Industrial and
Systems Engineering
Jane Russenberger, BS
Senior Director of
Genetics and Breeding
Contingent upon funding
15. Age
(Months)
Source of observation
Stored
51 data points
Number
2
• Socialization and training
• Puppy test
Behavior Checklist
(BCL) & Video
1500
4,8,13 • Puppies with raisers BCL & Video 1000
14-18 • Training induction test BCL & Video 1200
>14
• Preliminary blindfold test and trainer’s
observations –first 2 months or when rejected
• Final blindfold test and trainer’s observations
BCL 1200
Lots of Longitudinal Data
16. • Decreased ability to pay attention to cues
• Decreased fine motor control
• Over-reaction to stimuli (impacts
excitability, environmental soundness)
Unproductive State of Arousal
17. 17
Breeding- Weaning
$2K
Puppy Raising
$13 K
Watson
Help Guide Decisions
Outcome:
More Guide Dogs
from
$46,000/dog
Training
$31 K
Guides
& Breeders
Release very poor dogs earlier
Select best pups to raise as guide dogs
Best matches of dogs to volunteer raisers
More successful dogs
Guiding Eyes is using cognitive computing to make dramatic advances in the art and science of raising guide dogs. Using natural language processing (NLP) on structured and unstructured data, the system is trained to find correlations to success among myriad genetic, health, temperament and environmental factors—and continues to learn.
Dr. Tseng is supervising a group of students at San Jose State University in a project where Watson will undergo supervised learning
We have provided a lot of data about the dogs and puppy raisers
Working with a combination of structured and unstructured data.
Early stages
Dr. Tseng is supervising a group of students at San Jose State University in a project where Watson will undergo supervised learning
We have provided a lot of data about the dogs and puppy raisers
Working with a combination of structured and unstructured data.
Early stages