BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Time Symmetries and Neurosymbolic Learning for Dynamical Systems -
  Miles Cranmer\, Princeton University
DTSTART:20210111T163000Z
DTEND:20210111T173000Z
UID:TALK155452@talks.cam.ac.uk
CONTACT:Bingqing Cheng
DESCRIPTION:We develop a general approach to distill symbolic representati
 ons of a learned deep model by introducing strong inductive biases. We foc
 us on Graph Neural Networks (GNNs). The technique works as follows: we fir
 st encourage sparse latent representations when we train a GNN in a superv
 ised setting\, then we apply symbolic regression to components of the lear
 ned model to extract explicit physical relations. We find the correct know
 n equations\, including force laws and Hamiltonians\, can be extracted fro
 m the neural network. We then apply our method to a non-trivial cosmology 
 example: a detailed dark matter simulation\, and discover a new analytic f
 ormula which can predict the concentration of dark matter from the mass di
 stribution of nearby cosmic structures. The symbolic expressions extracted
  from the GNN using our technique also generalized to out-of-distribution 
 data better than the GNN itself. Our approach offers alternative direction
 s for interpreting neural networks and discovering novel physical principl
 es from the representations they learn.
LOCATION:virtual ZOOM meeting ID: 263 591 6003\, Passcode: 000042\, https:
 //us02web.zoom.us/j/2635916003?pwd=ZlBEQnRENGwxNmJGMENGMWxjak5nUT09
END:VEVENT
END:VCALENDAR