Thursday, January 25, 2018

10 Takeaways from AI World Conference in Boston in Dec

Last month I attended the 3 day AI World Conference in Boston with 2,300 others.  The event was cross-industry (digital advertising, health care, autonomous cars, security, etc.) with speakers from leading firms such as MIT, Goldman Sachs, Forrester, Phillips Health Care, Toyota, iRobot, etc.   Below are my key takeaways from my 15 pages of notes and conversations with dozens of people.

Abbreviations
AI: artificial intelligence
ML: machine learning
IoT: internet of things

Key Takeaways
  • The category labels (AI/ML/Big Data/IoT) continue to blur and merge, but “AI” is now commonly used as the larger termer umbrella term
    • There are different definitions of AI and there is no trend to consolidate 
    • Two speakers said, “Just substitute AI for ML in all places on my presentation”
    • Many “AI” case studies (health care, etc.) did not involve mechanical or robotic hardware
    • Speakers quoted stats about a large number of “AI” startups but these startups would have been traditionally classified software firms that use ML
    • Everything is called AI now: “AI for the Enterprise”  
    • “As long as the startup has one machine learning model, we call it a ‘AI’ startup“
  • The pace of technology change is relentless and increasing 
    • Many felt the pace of technology change continues to accelerate.  People within the AI/ML community stated that  “it is hard to keep up with AI/ML” and most feel this acceleration will continue.   
    • Companies and consumers will never be able to keep up which provides more opportunities for vendors to package all these capabilities.
  • This increase in technology remains explosive and drives automation
    • Goldman Sachs example
      • 2000: 600 cash equity traders, no software engineers
      • 2017: 2 cash equity traders and 200 software engineers
    • Health insurance open enrollment example  
      • call center=> web => AI
  • While AI and ML are hyped right now, it remains very, very early.  Internet firms (Google, Facebook, Twitter, Amazon) dominate current hiring and case studies
    • 80% [my estimate] of activity, advancement, and energy around AI/ML is from internet firms (Google, Facebook, Twitter, Amazon) which have massive data sets and massive teams and have hired the vast majority of ML talent at high wages
    • The rest of the companies are WAY behind these
  • The successful early applications and ones with the most promise in the next 3 years are those with NARROW feature set that supplement an existing process
    • Examples include
      • Uber: human driver 95% and uber app 5%.  More automation: multiple pickups (human couldn’t do this easily)
      • Social media: scan millions of images for NSFW but have human QA involved for exceptions
      • Internet advertising:  use ML to suggest banner ads but add “guardrail” rules for edge cases (e.g. don’t show ads for coffins when you believe someone has a death in the family)
      • Healthcare: analyze hundreds of MRI scans as second opinion for doctors
    • There were multiple names for this task augmentation with software and ML
      • “Augmented intelligence”
      • “Computer enabled service” 
      • “Co-bots"
  • Fully autonomous AI is decades or more away
    • Yes the computer can beat the human in a chess game, but the computer couldn’t find the table to sit down ;)
    • None of the speakers at the conference believe fully unassisted AI will be rolled out in the next 3 years (autonomous car panel thought it would be 3-10 years before autonomous cars appear in the market beyond current pilots)
    • The media hype is way ahead of realistic adoption
  • Data (and lots of it) are KING. Models are critical but of zero value without data
    • This was a recurring theme, especially from people who have been through any kind of real AI/ML project
    • The data collection challenge and cost is material in many industries.  For example, Electronic Health Records (EHR) is burdened by abbreviations, missing data, inconsistent usage, IT data silo’s and other challenges  
  •  Models can be imprecise and fooled (several examples)
    • Tape on stop sign fools autonomous car

    • Poodle or Ostrich? -  Neural network fooled by removal of several pixels (but human eye can still tell the difference)
    • Is a chihuahua or a muffin?  A labradoodle or fried chicken?
  • The data modeling processes are immature and rudimentary when compared to software engineering and the integration of these two teams is non-trivial
    • No methodology to develop, test, deploy, and document models exist
    • Better tools and processes are the second priority at MIT CSAIL, where they are advocating ModelDB
    • Many firms mentioned difficulty in integrating software engineers and data scientists
    • “Even at MIT, software engineering department is still too stovepiped from data scientist”
  • Neural networks are powerful but many shy away because there is no explainability
    • Neural networks remain very problematic because lack of reasons for prediction
    • Medical diagnostics (why does model say you have cancer?)  why buy this stock?
    • Find way to better explain output from neural networks is top research agenda at MIT per Sam Madden, co-chair of MIT CSAIL
Overall, speakers and attendees are very bullish on the value and competitive differentiation AI/ML bring to the market but the AI/ML market is still in its infancy and the press hype is ahead of realism.  Successful firms that automate more processes in the next 5 years will incorporate AL/ML as the cost of data acquisition and tools continue to decline.