BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Observation and Intervention Incentives in Causal Influence Diagra
 ms: Towards an Understanding of Powerful Machine Learning Systems - Tom Ev
 eritt (DeepMind)
DTSTART:20190125T110000Z
DTEND:20190125T120000Z
UID:TALK118648@talks.cam.ac.uk
CONTACT:Adrià Garriga Alonso
DESCRIPTION:As machine learning systems gain in capability and complexity\
 , understanding their incentives will become increasingly important. In th
 is paper\, we model their objectives and environment interaction in graphi
 cal models called influence diagrams. This allows us to answer two fundame
 ntal questions about the incentives of a machine learning system directly 
 from the graphical representation: (1) which nodes would the system like t
 o observe in addition to its observations\, and (2) which nodes would the 
 system like to control in addition to its actions? The answers tell us whi
 ch information and influence points need extra protection\, and have appli
 cations to fairness and reward tampering. For example\, we may want a clas
 sifier for job applications to not use the race of the candidate\, and a r
 einforcement learning agent not to take  direct control of its reward mech
 anism. Different algorithms and training paradigms can lead to different i
 nfluence diagrams\, so our results can help designing algorithms with  les
 s problematic observation and intervention incentives.
LOCATION:Engineering Department\, LR5 (in front of library\, 1st floor)
END:VEVENT
END:VCALENDAR