Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Write/Speak/Code: Challenging Algorithm Development

Write/Speak/Code: Challenging Algorithm Development

Liz Rush

June 16, 2016
Tweet

More Decks by Liz Rush

Other Decks in Programming

Transcript

  1. ALGORITHM DEVELOPMENT WHERE DO YOU FIND ALGORITHMS? ▸ Tools: recommenders,

    suggested friends, or auto-tagging ▸ Business Intelligence: customer behavior prediction ▸ AI: email assistants, customer service chat bots ▸ High stakes autonomous systems: self-driving cars, drones ▸ Basically everywhere
  2. WHAT LURKS WITHIN THESE ‘BLACK BOXES’? HOW CAN THEY BE

    UNDERSTOOD, GOVERNED AND MADE ACCOUNTABLE? THESE ARE CHALLENGING BUT SIGNIFICANT POLITICAL QUESTIONS, WHICH REQUIRE URGENT PUBLIC DEBATE. Dries Buytaert, creator of Drupal ALGORITHM DEVELOPMENT
  3. WHY YOU SHOULD CARE EXAMPLES OF BIAS IN ALGORITHMIC DESIGN

    ▸ Algorithm design for office temperature optimization ▸ Facebook’s 2014 “emotional contagion” mood manipulation ▸ Search: ▸ Men are more likely to be shown Google Ads for high paying jobs ▸ Ads for arrest records more likely to show up on searches for distinctively Black names or a historically Black fraternities ▸ Learned autocomplete ▸ AI assistants returning jokes when users complain of assault (e.g. Siri)
  4. WHY YOU SHOULD CARE POTENTIAL BIASES ▸ According to the

    American Institute for Behavioral Research & Technology, information displayed in Google’s search engines could shift voting preferences for undecided voters by 20% or more ▸ Voting Machines ▸ Medical Devices ▸ Algorithms to determine early release from prison ▸ Breathalyzers - Algorithms behind breathalyzers have been requested in courts but denied access
  5. ALGORITHMIC BIAS CLASSIFICATION ACCURACY AS SOURCE OF UNFAIRNESS ▸ Unfairness

    goes far beyond insensitivity & insults to arbitrarily limit people’s opportunities ▸ Most ML objective functions create models accurate for the majority class at the expense of the protected class ▸ For minority populations, the number of training samples is dwarfed by the majority ▸ Accuracy-Fairness tradeoff
  6. DISCRIMINATION IS AN EMERGENT PROPERTY OF ANY LEARNING ALGORITHM
 Delip

    Rao, machine learning expert WHO IS THIS DESIGNED FOR?
  7. WHO IS THIS DESIGNED FOR? HOW DO WE GET TO

    FAIRNESS? ▸ No current model for algorithmic fairness ▸ We don’t even have one “best” conception of what fairness actually is! ▸ One way to characterize fairness is to ensure both majority and the protected population have similar outcomes. ▸ We know we must involve as many people as possible in design & discussion of algorithms
  8. SOLVE FOR SERENDIPITY FALSIFIABILITY OVER OPTIMIZATION ▸ Obsession with optimization

    leads to missed opportunities ▸ Netflix’s “Recommended for you” as example of optimization leading to tunnel vision ▸ User centered design values > pure optimization ▸ If the algorithm is always trying to disprove its own model, we avoid over-fitting ▸ Allows for “serendipity”
  9. MOST ALGORITHMS ARE PROPRIETARY & CLOSED- SOURCE “OPEN-SOURCE” ALGORITHM APIS

    & “OPEN AI” ARE OFTEN ONLY PARTIALLY OPEN-SOURCE
  10. Dries Buytaert, creator of Drupal ALGORITHMIC ACCOUNTABILITY WE NEED TO

    KNOW WHICH DATA IS CAPTURED, HOW THAT DATA IS USED, BUT ALSO HOW THESE ALGORITHMS WORK…IT WOULD BE GOOD IF SOMEBODY COULD AUDIT THESE ALGORITHMS TO BE SURE THERE ISN’T BIAS BUILT INTO THEM —EITHER ON PURPOSE OR ACCIDENT
  11. ALGORITHMIC ACCOUNTABILITY OPEN MARKETPLACES ▸ Marketplaces allow us to buy,

    sell, and rent algorithms now ▸ Allows non-algorithm developers to make business choices about which algorithms to use ▸ Follows “free market” principles: ▸ The best algorithms will become the most popular ▸ People will seek out algorithms specific for their needs ▸ Creates space for two-way communication between creators and users
  12. ALGORITHM MARKETPLACES ARE SIMILAR TO THE MOBILE APP STORES THAT

    CREATED THE ‘APP ECONOMY’. Alexander Linden, research director at Gartner THE ALGORITHM ECONOMY
  13. THE ALGORITHM ECONOMY WHAT ARE THE FEATURES OF ALGORITHM MARKETPLACES?

    ▸ Similar to micro service & SOA architecture ▸ More cutting edge algorithms on the market; fewer stuck in academia and lower barrier of entry ▸ Standardization through reuse and chaining ▸ Commercialization incentivizes accuracy & validation ▸ Follows “free market” principles: ▸ Not necessarily concerned with ethics ▸ People can choose to use biased algorithms
  14. WE MAY NEED A FEDERAL ROBOTICS COMMISSION TO HELP OTHER

    AGENCIES, COURTS, AND STATE AND FEDERAL LAWMAKERS UNDERSTAND THE TECHNOLOGY WELL ENOUGH TO MAKE POLICY. Ryan Calo, cyberlaw expert at Washington University ALGORITHMIC ACCOUNTABILITY
  15. ALGORITHMIC ACCOUNTABILITY AN F.D.A. FOR ALGORITHMS ▸ A federal agency

    to oversee companies’ algorithms ▸ Proposed by many in the Open Web movement ▸ 2014 White House report on Big Data raised concerns about privacy and fairness ▸ India recently banned “Free Basics” provided by internet.org for violating the essential rules of net neutrality
  16. ALGORITHMIC ACCOUNTABILITY OMBUDSMANSHIP - PRIVATE GOVERNANCE ▸ Lessons from newspapers:

    ▸ It can be in a business’s best interest to have oversight ▸ Business lead the push towards corporate responsibility ▸ Established the idea of “public interest” in journalism ▸ Led to legal rules accordingly ▸ Consumers began demanding more rigorous oversight
  17. TRANSPARENCY DOESN'T MAGICALLY REDUCE CHEATING OR IMPROVE SOFTWARE QUALITY, AS

    ANYONE WHO USES OPEN-SOURCE SOFTWARE KNOWS. IT'S ONLY THE FIRST STEP. Bruce Schneier, Schneier on Security ALGORITHMIC ACCOUNTABILITY
  18. ALGORITHMIC ACCOUNTABILITY WHOSE OMBUDSMANSHIP, THOUGH? ▸ Google has an artificial

    intelligence ethics board, founded after buying DeepMind ▸ They refuse to name who is on the board ▸ Not transparent about what they do ▸ Other AI startups are more inclined to share who is on their ethics boards, such as Lucid.AI, which has made their board public ▸ If we don’t know who is on the ethics oversight boards, how do we know they are ethical?
  19. AS SYSTEM DESIGNERS, WE HAVE A RESPONSIBILITY (AND OPPORTUNITY) TO

    DESIGN SYSTEMS WITH STRONGER VALUES. THEY MAY NOT CHANGE US (WE ARE OLD), BUT OUR CHILDREN WILL SEE THE VALUES IN THESE SYSTEMS AS NORMAL. THAT IS BOTH SCARY AND EXCITING. Buster Benson on Eric Meyer’s XOXO 2015 Conference Talk CHOICE
  20. RESOURCES FURTHER READING ▸ Jamis Buck’s “Algorithm is Not a

    Four-Letter Word” for RubyConf 2011 ▸ Rachel Shadoan’s “Reasoning About Opaque Algorithms” in The Recomplier ▸ Cathy O’Neil’s “Weapons of Math Destruction: How Big Data Increases Inequality & Threatens Democracy” ▸ Eli Pariser’s “The Filter Bubble” ▸ “The Secret Rules of Modern Living: Algortihms” on Netflix
  21. ▸ USE OPEN SOURCE ALGORITHMS & REPORT BIAS ▸ WRITE

    ABOUT WHAT YOU SEE IN THE WILD ▸ EXPLAIN TO NON-TECHNICAL FRIENDS ABOUT THE SUBJECTIVITY IN ALGORITHMS ▸ ALWAYS KEEP TALKING! WHAT CAN YOU DO?