Tuesday, December 12, 2017

Notes Day 1: AI World, Dec 11-13, 2017 Boston

My Day 1 Notes AI World Conference Dec 11-13, 2017 Boston
2200 attendees

Monday Morning

Srividya Sridharan, VP Research Director, Forrester
Michelle Goetz, Forrester Analyst
Definition: AI is a system for machine to interact, think or mimic human intelligence or engagement
·      Historically, humans had to adjust to machines.  Trends now is for machines to adjust to humans (e.g. voice commands)
·      Case study: AON Hewitt open enrollment,
o   Previously: seasonal workers, poor
o   New: trained AI and self service
·      Retail example
o   Traditional BI world: what should I do next?
o   AI Driven: What objective will I reach?
§  Tell “machine” what you want to do and machine figures out to
·      Per Forrester, AI move from mechanical mindset to sentient mindset
·      Must start with purpose and DESIGN. Design thinking is the preferred approach, not just another tool.  Avoid unused data lakes, etc.
o   Art of design
o   Process design
o   Beyond analytics
o   True Collaboration
o   Agile development – focus on outcome and quality

Mr. William Mark, President SRI International
Dany De Grave, Director, Sanofi Pasteur

·      Toyota: what if you had an emotional connection with the car.  What if the car care about you
·      William Mark: trust is an important element of the solution.  Explain why things are happening

David Kiron, MIT Sloan Management Review – Adoption of AI in business: Opportunities and Challenges

·      More than 2,000 AI startups
·      China and Canada, Israel top AI startups after the US
·      Only 10,000 people in the world can do serious AI [i.e. ML/data science]
o   Google hired 40 people from CMU data science dept
o   7 figure comp for AI people
o   average salary $350K at Google deep learning team
o   academia simply can’t manufacturer data scientists fast enough
·      Shortage of AI is serious industry problem
·      No amount of algorithmic sophistication will overcome a lack of data.
·      more mature AI-experience teamed and companies understand the need for training models and software development
·      business drivers
o   competitive advantage 84%
o   move into new business
o   fear: competitors are doing this

Kjell Carlsson, Senior Analyst, Forrester, PhD –AI for Super Human Insight
·      commonalities of high profile in AI use cases
o   self driving cars.  Google self driving cars now 10 years
o   virtual agents (Siri).  Siri 10 years old. But still years away from CSR other than basic questions
·      a long time to value
o   over promise. Don’t meet expectations
·      Classify applications in 3 categories
o   AI for subhuman automation
§  Hard time with variety or context
§  Can’t fill gaps
§  Can’t reason or explain
§  Can’t instill confidence
o   AI for non human automation
§  Performs things that we feel bad asking people to do
§  Reset passwords, data entry, moving things around a warehouse, connecting with the right person
§  ROI is there, but will run out
§  Drivers less cars                       Driver assistance GPS, collision warning, blind spot warning, land drifing, ride sharing
o   Medical care
§  Medical imaging: skin cancer, breast cancer, ovarian cancer, lung cancer
o   AI for Super Human insights
§  Common features:
·      Data beyond what any human can do
·      Analysis: more powerful than any mind
§  Complexity patterns too complex
§  Speed: faster than a speeding bullet
§  Use
·      Applied stats, traditional ML, deep learning, text analysis, NLP, image and video analytics, speech analytics, AI APIs
§  Examples
·      Semi-truck assistance (remote human)
·      Physician recommendation systems  [but how enough data?]
·      (who wins chess tournaments?   AI only or AI with Human Machine team

Future of work: JP Gownder, VP Forrester
·      transition from human be like computer to computer be like human
·      business processes are algorithms
·      example
o   1960 Budget director: manual budget calculation
o   2017  Budget director: more strategic, automation
·      “Co bots”
·      Goldman Sachs
o   2000: 600 cash equity traders, no software engineers
o   2017: 2 cash equity traders and 200 software engineers
·      prediction: mixed human/digital workers
o   AI infused physical robots
o   Human machine teaming

Don Schuerman, CTO, VP Product Marketing, Pegasystems, Inc.

·      “Machine reasoning”
o   we teach the computer the rules
o   train the computer to discover the rules
o   neural networks. Can’t give reasons or explain
·      had machine learning model to learn painter Bob Ross
·      good today
o   tag friends on facebook automatically
·      Opaque AI  - can’t tell reasons
·      Transparent AI – can show reasons about how it came to a solution
·      Facebook “likes” can predict things
o   10 likes; 70% gender, openness,
o   70 likes better than you friends
o   150
·      NBA – Next Best Action for customer engagement
·      Decision Management
o   Combination of rules, ML that is transparent and testable for bias
o   ML gated with business rules

Michael Facemire, VP Forrester – the common thread across the future of experiences

#1 question from customers, how to do I build a chatbot, but business problem definition and overall design approach

State of AI in Medicine
Data Scientist in LA Children’s Hospital
Phillips Healthcare

Philips Health Care
Smart Exam of MRI, MR
IntelliSpace Portal for radiologist
Illumeo – serves patient briefing

Day Zero Diagntoics – Miriam Huntley, PhD
·      Why now
o   1. Lots of data now available
o   2. Model maturity
·      face recognition
o   1977: small data sets, features on map
o   2017: facebook millions of parameters,
·      ML in healthcare
o   Medical imaging
§  Give you an image and predict cancer, tumor cells, etc.
§  LOTS of high quality data (but does take time to procure)
§  Uses “deep learning” neural networks. mature
o   EHR (electronic health record)
§  Given a HER, predict appropriate treatment, drug
§  Data is available but poor quality (unstructured text, abbreviations, spelling mistakes)
§  Models are NLP, topic models,
·      ML methods still in development with focus success (e.g Google Translate)
o   Genomics
§  Task: given a patient genomic data, predict disease risk for baby or cancer presence
§  Data is the problem. Cost of sequencing was $100M but now $1000 per person.
§  Models are simplistic
·      Bayesian
·      Day Zero Diagnostic
·      Promising future applications: new

·      Yin Aphinyanaphongs, PhD, MD
·      NYU – why models not used in health care

·      Why models not used
                                    poor coupling of data
·      80/20 rule for “last mile”.   Last 20% of model accuracy can take 80% of effort and time.  Depends
·      Health Records
o   Dominated by 3 vendors (Epic, etc.)
o   Built around billing, not optimized for
o   Doctor 28 hour shift make 300 decisions
§  Visualization is critical for doctors, rather than another warning system (is this really the dose you want?)
o   Hard to find shared common language between medical field and IT/software industry
§  If you want doctors/clinicians involved, you need to dumb down the tech
§  Keep information very simple
§  Max use of visualization
o   Promising
§  Imaging, replace/augment image interpretation
§  HER untapped potential
o   Challenges
§  What to do with the prediction?  Patient x has 25% chance of x chance.  What does doctor do with this data?
§  Privacy. Can’t always see data always for heart rate for last 30 min
§  Trust. What if model is wrong?  Every tesla accident is on CNN.
§  Algrotith conversion Chicago business school
·      People are extremely critical of machine models
·      Claes Gustafsson, co founder ATUM
·      Ravi NTT Innovation Institute
o   Trends in healthcare
§  Be healthy
§  Personalized care
§  Consumer driven
§  Post patient care/ beyond the pill
o   Focus on two problems
§  Medical emergencies
·      130M medical emergencies yearly in the US
·      15M strokes per year
o   device detectors (motion sensors, wearables,
o   score
§  1. Gait
§  2. Record voice – look for search
§  3. MS Kinect:  watch how you drop your hand after holding for 5 seconds
o   results
§  reduced 96% false positives
§  adverse drug reactions (ADR)
·      smart pill box
o   AI and Pharma
§  Dany DeGrave, Sanofi
·      Vaccines. Take 10-20 years to work
·      Data project
o   Results??
§  Had independent team also evaluate. Validate drug success?
·      lessons
o   Who is on the team is important
§  Need people who are curious and aware of bias
§  Wrong people can ruin the project
o   It takes time
§  Peter Henstock, PHD,  Pfizer
·      Computers can
o   Recognized words
o   Write
o   Images
·      Industries are rapidly changing (transportation, ag, etc.)
·      Why not pharma?
·      150 data analyst/statistians. This FTE is not increasing
·      why not big data not in pharma
o   silo’d data
§  if you can’t do basic analytics on your data, then you can’t do AI/ML
o   pharma is not tech driven
Marteh Ghessmie, PhD, MIT Machine Learning and health.  Works at Googe Verily
·      only 10%-15% of treatments in ICU have been based on RCT, the gold standard approach for trials due to cost
·      many data challenges
o   HER – now used at 80% of US hospitals has promise but many abbreviations
o   Missing data and time frames
§  Inconsistent time measurements (measure every 30 min or every hour)
§  Measured but not recorded
§  Unknown.  Medicine taken but not recorded

VC Panel:
Glasswing – Rudina
Autonomy Ventures – Michael
Accenture –
·      Rudina
o   Excited about “narrow AI”
o   AI is a part of the total solution
o   Very focused on problem being solved
·      Michael
o   East cost VC firms chase revenue as validation. West cost firms chase concepts
o   His firm happy with exists