BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Structured recognition for generative models with explaining away 
 - Changmin Yu\, Gatsby Computational Neuroscience Unit\, UCL\, London\, UK
DTSTART:20231002T123000Z
DTEND:20231002T133000Z
UID:TALK205996@talks.cam.ac.uk
CONTACT:Samuel Eckmann
DESCRIPTION:A key goal of unsupervised learning is to go beyond density es
 timation and sample generation to reveal the structure inherent within obs
 erved data.  Such structure can be expressed in the pattern of interaction
 s between explanatory latent variables captured through a probabilistic gr
 aphical model.  Although the learning of structured graphical models has a
  long history\, much recent work in unsupervised modelling has instead emp
 hasised flexible deep-network-based generation\, either transforming indep
 endent latent generators to model complex data or assuming that distinct o
 bserved variables are derived from different latent nodes.  Here\, we exte
 nd the output of amortised variational inference to incorporate structured
  factors over multiple variables\, able to capture the observation-induced
  posterior dependence between latents that results from "explaining away''
  and thus allow complex observations to depend on multiple nodes of a stru
 ctured graph.  We show that appropriately parameterised factors can be com
 bined efficiently with variational message passing in elaborate graphical 
 structures. We instantiate the framework based on Gaussian Process Factor 
 Analysis models\, and empirically evaluate its improvement over existing m
 ethods on synthetic data with known generative processes. We then fit the 
 structured model to high-dimensional neural spiking time-series from the h
 ippocampus of freely moving rodents\, demonstrating that the model identif
 ies latent signals that correlate with behavioural covariates.
LOCATION:CBL Seminar Room\, Engineering Department\, 4th floor Baker build
 ing
END:VEVENT
END:VCALENDAR