BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:AI+Pizza - Microsoft Research/University of Cambridge
DTSTART:20180525T163000Z
DTEND:20180525T180000Z
UID:TALK106405@talks.cam.ac.uk
CONTACT:Microsoft Research Cambridge Talks Admins
DESCRIPTION:Speaker 1: Brooks Paige\nTitle: Predicting Electron Paths\nAbs
 tract: I will be talking about our recent NIPS submission on modeling chem
 ical reactions by predicting electron paths. Chemical reactions can be des
 cribed as the stepwise redistribution of electrons in molecules. As such\,
  reactions are often depicted using “arrow-pushing” diagrams which sho
 w this movement as a sequence of arrows. We propose an electron path predi
 ction model to learn these sequences directly from data. Instead of predic
 ting product molecules directly from reactant molecules in one shot\, lear
 ning a model of electron movement has the benefits of (a) being easy for c
 hemists to interpret\, (b) incorporating constraints of chemistry\, such a
 s balanced atom counts before and after the reaction\, and (c) naturally e
 ncoding the sparsity of chemical reactions\, which usually involve only a 
 small number of atoms in the reactants. We design a method to extract appr
 oximate reaction paths from any dataset of reaction SMILES strings. Furthe
 rmore\, we show that the model recovers a basic knowledge of chemistry wit
 hout being explicitly trained to do so. Joint work with John Bradshaw\, Ma
 tt Kusner\, Marwin Segler\, and José Miguel Hernández-Lobato.\n\nSpeaker
  2: Alexander Gaunt\nTitle: Constrained Graph Variational Autoencoders for
  Molecule Design\nAbstract: Graphs are ubiquitous data structures for repr
 esenting interactions between entities.\nWith an emphasis on the use of gr
 aphs to represent chemical molecules\, we explore\nthe task of learning to
  generate graphs that conform to a distribution\nobserved in training data
 . We propose a variational autoencoder\nmodel in which both encoder and de
 coder are graph-structured.\nOur decoder assumes a sequential ordering of 
 graph extension steps and we\ndiscuss and analyze design choices that miti
 gate the potential downsides of this linearization.\nExperiments compare o
 ur approach with a wide range of baselines on the molecule\ngeneration tas
 k and show that our method is more successful at matching the statistics\n
 of the original dataset on semantically important metrics. Furthermore\, w
 e show\nthat by using appropriate shaping of the latent space\, our model 
 allows us to\ndesign molecules that are (locally) optimal in desired prope
 rties.\n\n
LOCATION:Auditorium\, Microsoft Research Ltd\, 21 Station Road\, Cambridge
 \, CB1 2FB
END:VEVENT
END:VCALENDAR