2. IBM is at the intersection of science and technology
Approximately
100 cognitive
health patents
filed in 2016
More than 7,000
employees
globally
We partner with industry
leaders to transform life
sciences.
IBM Watson Health Life-Sciences offerings
Drug development value chain
Market assessment
and molecule
development
(preclinical)
Clinical trial
protocol
development
Site selection
and patient
recruitment
Study conduct,
data collection
and operations
Study analysis,
clinical study
report and
submission
Postmarketing
and real-world evidence
studies (postmarket
requirements and
commitments)
Safety and
pharmaco-
vigilance
Research CommercializationDevelopment
Real World
Evidence
Watson for
Patient Safety
Watson™ Drug
Discovery
IBM Clinical
Development
Data assets:
Truven and Explorys®
software
From molecule …. to market…
More than 1500 Clinical
Trials conducted on our
technology
http://ibm.biz/wdd-tria
4. 4
2010 2025
160 Zettabytes by 2025*
We are here
Computing power
Data growth Advances in neural networks,
machine learning and deep learning
Cloud
Data-centric
Systems
End-to-end
Security
Extreme
Scalability
*Source IDC. IBM projections based on analyst report
AI: Why now?
80% Unstructured
5. The Evolution of AI
We are here
Broad General
1980 2015 2050? Beyond
Narrow
6. “I feel so strongly that deep and simple is far more
essential than shallow and complex.”
-- Mr. Rogers
Guiding Principle:
7. Ten Corollaries to Mr. Rogers’ Law
1.Interpolation is preferable to Extrapolation
2.Confidence of “facts” must be accurately estimated
3.Cognitive Reasoning is useless without explanation
4.Structure implies behavior
5.The past is validation for the future
6.Unlikelihood is a measure of interestingness
7.Don’t look for an answer, look for a question
8.Use what you know as a map of the unknown
9.To know what you need to know, first know what you know.
10.Persistence is a virtue
8. Interpolation vs Extrapolation
Use what you know to fill in the gaps of what you do not know… far
more is implied than stated.
Implication rather than Prediction
9. Confidence Scoring – how to know what to believe
Two Factors Determine the Likelihood of a Fact being True
• Frequency of Occurrence (higher frequency -> more likely to be true)
• Consistency, measured by collaborative filtering (higher matrix factorization score ->
more likely to be true).
• At the limits:
• Infinite support = 100% confidence
• 1.0 MF = 100% confidence
• 0.0 MF and 1 support = lowest confidence
• P = aMFd
– b/Se
+ c
• Need to determine a, b, c, d and e experimentally.
• Process: Sample facts with different levels of Su and MF and determine the observed
P value.
14. Don’t look for an Answer…
…. Look for a (better) Question
Don’t expect AI to be an oracle. Think of it as a collaborator. A really annoying one who
keeps saying, “What about…X?”