BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Machine Learning Industry &amp\; Academic Perspectives - Prof Migu
 el Hernandez Lobato - University of Cambridge &amp\; Ben Pellegrini - Inte
 lligens Ltd.
DTSTART:20240109T150000Z
DTEND:20240109T160000Z
UID:TALK210148@talks.cam.ac.uk
CONTACT:Alex Wilby
DESCRIPTION:*Topics:*\n\nScaling up Gaussian processes\nwith stochastic gr
 adient descent:\n\nGaussian processes are a powerful framework for\nquanti
 fying uncertainty and for sequential decision-making\nbut are limited by t
 he requirement of solving linear systems.\nIn general\, this has a cubic c
 ost in dataset size and is\nsensitive to conditioning. We explore stochast
 ic gradient\nalgorithms as a computationally efficient method of\napproxim
 ately solving these linear systems. Experimentally\,\nstochastic gradient 
 descent achieves state-of-the-art\nperformance on large-scale regression t
 asks. Its uncertainty\nestimates match the performance of significantly mo
 re\nexpensive baselines on a large-scale Bayesian optimization\ntask. On a
  molecular binding affinity prediction task\, our\nmethod places Gaussian 
 processes with a Tanimoto kernel\non par with state-of-the-art graph neura
 l networks.\n\nDeploying machine learning in\nchemical industry:\n\nHow do
  we separate AI hype from reality? This talk focuses\non the use of machin
 e learning (ML) for data analysis and\noptimising experiments\, products\,
  and processes in\nresearch-intensive industries. ML must be seen as a too
 l\nthat complements rather than replaces human expertise.\nThis means thin
 king about issues such as how to formulate\nthe right questions to ask the
  ML\, time constraints on its use\,\nhow to pilot it successfully in deplo
 yments\, and investment\nin optimising user experience. The discussion wil
 l touch on\nthe in-house versus external software debate\, suggesting a\nb
 alanced approach for optimal results.Case studies will\nshow how\, once so
 me or all of these factors are considered\,\nML has been deployed to reduc
 ed experimental workloads\n(typically by 50-80%)\, enhance data insights\,
  and enable the\ndesign of improved products and processes\n\n
LOCATION:Pfizer Lecture Theatre - Yusuf Hamied Department of Chemistry
END:VEVENT
END:VCALENDAR