BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Noisy\, sparse\, nonlinear: Navigating the Bermuda Triangle of phy
 sical inference with deep filtering - Carl Poelking
DTSTART:20200511T160000Z
DTEND:20200511T163000Z
UID:TALK141877@talks.cam.ac.uk
CONTACT:Bingqing Cheng
DESCRIPTION:Capturing the microscopic interactions that determine molecula
 r\nreactivity poses a challenge across the physical sciences. Even a basic
 \nunderstanding of the underlying reaction mechanisms can substantially\na
 ccelerate materials and compound design\, including the development of\nne
 w catalysts or drugs. Given the difficulties routinely faced by both\nexpe
 rimental and theoretical investigations that aim to improve our\nmechanist
 ic understanding of a reaction\, recent advances have focused on\ndata-dri
 ven routes to derive structure-property relationships directly\nfrom high-
 throughput screens. However\, even these high-quality\,\nhigh-volume data 
 are noisy\, sparse and biased - placing them in a regime\nwhere machine-le
 arning is extremely challenging. Here we show that a\nstatistical approach
  based on deep filtering of nonlinear feature\nnetworks results in physico
 chemical models that are more robust\,\ntransparent and generalize better 
 than standard machine-learning\narchitectures. Using diligent descriptor d
 esign and data\npost-processing\, we exemplify the approach using both lit
 erature and\nfresh data on asymmetric catalytic hydrogenation\, Palladium-
 catalyzed\ncross-coupling reactions\, and drug-drug synergy. We illustrate
  how the\nsparse models uncovered by the filtering help us formulate\nphys
 icochemical reaction "pharmacophores"\, investigate experimental bias\nand
  derive strategies for mechanism detection and classification.\n\n
LOCATION:virtual ZOOM meeting ID: 263 591 6003\, https://zoom.us/j/2635916
 003
END:VEVENT
END:VCALENDAR