BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning curve prediction for AutoML - Andy Lin
DTSTART:20241204T110000Z
DTEND:20241204T123000Z
UID:TALK225259@talks.cam.ac.uk
CONTACT:120952
DESCRIPTION:Automated machine learning (AutoML) aims to automate the proce
 ss of selecting hyper-parameters for machine learning models\, such as lea
 rning rate\, batch size\, or layer width. To this end\, machine learning m
 odels are trained with different hyper-parameter configurations\, their fi
 nal performance is recorded\, and new candidate configurations are selecte
 d via Bayesian optimisation. The latter typically constructs a probabilist
 ic surrogate of final model performances as a function of hyper-parameter 
 configurations. However\, individual training runs are usually subject to 
 intermediate evaluations\, which produce learning curves in addition to th
 eir final performance. These learning curves could be leveraged to 1) save
  resources by stopping unpromising runs early\, and 2) improve the probabi
 listic surrogate to select better candidate configurations. This reading g
 roup will review the literature on building scalable probabilistic surroga
 te models of such learning curves\, discussing approaches using Gaussian p
 rocesses\, power laws\, Bayesian neural networks\, and Transformers.
LOCATION:Cambridge University Engineering Department\, CBL Seminar room BE
 4-38.
END:VEVENT
END:VCALENDAR