The document tells the story of Jean-Sebastian Gagnon from birth through his childhood and education into adulthood. As a child, he was inspired by books from his great aunt and showed early interest in music, science and questioning what makes humans special. He later studied computer science and became a startup founder working on applications of artificial intelligence in areas like healthcare and security. The document discusses the history and developments of AI technologies over time and how JS continues to promote responsible advancement of AI through education.
64. Bolukbasi, Chang, Zou, Saligrama, Kalai, 2016
Man : King :: Woman : Queen
Man : Computer Programmer :: Woman : Homemaker
Black Male : Assaulted :: White Male: Entitled To
Inherent Bias in Word Embeddings
The first neural network. enthusiasm was destroyed by the 1980s book by Marvin Minsky and Seymour Papert that proved mathematical limitations to neural networks. This led to an AI winter.
- 1952 “Theseus” mouse who learns how to go through a maze
Lowers the barrier to entry for use in enterprises
Orange means new since 2015!
PyMC is helps with bayesian - Tensorflow from Google
Gensim for unsupervised learning and LDA
- it’s actually 10 friends in 14 days, but who cares
word embeddings map words into vectors and then compress them down
neat things about these vectors are that directions in space seem to encode semantic meaning; and difference vectors encode analogies (e.g., woman-man; queen-king).
Image from Colah’s blog with embeddings of all Wikipedia
debiasing method was to send pairs out to M-turk to get feedback on which qualify as biased in mind of beholder
and then collapse all above the line down to vertical to remove associations with gender pairs