Machine Learning the Future

This is class 1pm-2:15pm Mondays in Spring 2017 at Cornell Tech and Cornell about the frontier of machine learning. Each class is lecture and discussion around a chosen topic. I would like each student to leave with an understanding of the topic deep enough to think about the next step.

The class will cover topics broadly related to reinforcement learning including contextual bandits, imitation learning, and exploration as well as other select topics drawn from online learning, optimization, boosting, parallelization, logarithmic prediction, active learning, representation. Logistically, we'll use zoom in addition to a video link.



January 30: cancelled (My voice is not working)
Feburary 6: ML the Future slides, Generalization slides with source, background reading, and a recording.
February 13: cancelled (travel)
Feburary 20: No class (February break)
Feburary 27: Online Linear Learning slides with source recording papers: Importance Invariant, Adagrad, Normalized Dataset: RCV1 CCAT-or-not
March 6: Contextual Bandit Eval and Optimization slides with source recording papers: Double robust policies Dataset: RCV1 CCAT-or-not in contextual bandit format (uniform exploration)
March 13: Contextual Bandit Exploration slides with source and recording, Papers Epoch Greedy, Cover, Bootstrap Dataset: RCV1 CCAT-or-not in multiclass format
March 20: Nonstationary evaluation slides with source Decision Service slides Contextual Decision Process slides recording Nonstationary evaluation, Decision Service, and Bellman Rank papers Decision Service site
March 27: Joint prediction as learning to search recording DAgger paper AggreVaTe paper
April 3: no class(Spring break)
April 10: Joint prediction results, analysis, programming slides with source recording LOLS paper Credit Assignment Compiler paper
April 17: Cancelled (healing up)
April 24: Logarithmic time prediction with source recording covering Error-Correcting Tournaments, Label embedding trees Dynamic Trees I, and Dynamic Trees II. Not covered: Static contextual bandits, Dynamic log time class probability Efficient Multilabel
May 1: Active Learning with source a recording Iwal papers search papers Cost sensitive active learning
May 8: Parallel Learning with source a recording allreduce paper distbelief paper 1-bit SGD Ring allreduce