Sannolikhet, statistik och kombinatorik: Diffusion interpolation in auto encoders

  • Datum: –11.15
  • Plats: Ångströmlaboratoriet 64119
  • Föreläsare: Henrik Hult
  • Arrangör: Matematiska institutionen
  • Kontaktperson: Xing Shi Cai
  • Seminarium

Auto encoding models have been extensively studied in computer science. They provide a framework for simulation, as well as for analysing feature learning. Furthermore, they are efficient in performing interpolations between data-points in semantically meaningful ways. We introduce a method for generating stochastic sequences from auto encoders trained on flattened sequences (e.g video sample from auto encoders trained to generate a video frame); as well as a canonical, dimension independent method for generating stochastic interpolations. The methods are based on diffusion bridges and we use the variational auto encoder model for demonstration purposes.  This is joint work with Carl Ringqvist, Judith Butepage, and Hedvig Kjellström.