6 Meetup - USF

6.1 Analysing & Preventing Unconscious Bias in Machine Learning. Rachel Thomas - 2018-10-19

https://www.meetup.com/USF-Seminar-Series-in-Data-Science/events/254217548/ Video: facebook.com/usfca.msds/

Abstract: Increasingly AI is finding its way into nearly every product we use (everything from photo sharing apps to criminal justice decision algorithms), but often various types of bias are buried in the underlying data and models. This can have a damaging impact on both individuals and society. Through the lens of 3 case studies, we will look at how to diagnose bias, identify some sources, and some steps towards addressing it.

  • Check out gendershades.org
    • Good example of use of data & diversity at varying levels of technical detail
  • Word Embeddings
    • Word2Vec - google library of word embedings
    • Stanford has a similar libraries
    • Rachel Thomas - word embeddings youtube
    • github: fastai/word-embeddings-workshop

  • ML can amplify bias

  • Compass software:
    • Determining who has to post bail
    • sentencing
    • parole
  • Problems:
    • Runaway feedback loops (predictive policing, etc.)
    • Ethical variables to include - recedivism algorithms?
  • Solutions:
  • Questions:
    • Bias in data
    • Code auditable? Open source?
    • Error rates for different sub-groups
    • Accuracy of simple, rule-based alternative?
    • Appeals process for mistakes?
    • How diverse is the team building it?
      • Diverse teams perform better
      • Believing you are meritocratic INCREASES bias
    • How do we address more nuanced biases once low hanging fruit are addressed?
      • Can be an interesting conversation, but don’t let perfect be the enemy of the good.

6.2 USF Intro session - 2018-10-19

  • FAQ
  • 12 Month accelerated program
    • 6 7-week modules & 1 2-week intersession course
    • 7/8/19 - 6/28/19
    • 60-80 hrs/wk
    • Practicum:
      • Similar to an internship, but with faculty mentors
      • Want to ensure good probems / work mentors
  • Highly recommend applications by 12/5 - better scholarship opportunities, ore thorough review of applications
  • Pre-reqs: linear algebra @ accredited university
    • Don’t have to have completed prereqs, but need a defined plan.
  • Send info ?s to info@datascience
  • Faculty interview:
    • Programming
    • inferential statistics
    • linear algebra
  • Personal statement: only 2 pages (won’t even read page 3)
    • show genuine interest in why this particular program
    • where below average, address the issues
      • want self-awareness - everyone struggles in the program - need to be able to self-evaluate
  • Letters of rec
    • Strong letter of rec with detail of why fit for this program
    • Work experience = 1 academic, 1 professional preferred
    • Both can be from work if no reasonably current academic references