Interactive companion to the paper · runs in your browser

Neural noise enables accurate internal simulation of rare events

A predictive network learns a world in which some events are rare. Add the right amount of noise to its internal replay and its simulation of those rare events becomes accurate. Turn the knob and watch.

Heng Zhang1,*, Pawel Herman2,#, Zenas C. Chao1,*

1WPI-IRCN, UTIAS, The University of Tokyo, Japan2KTH Royal Institute of Technology, Sweden

*Corresponding author#Co-senior author

Ground truth
Target
Simulation
Model
Rare-event rate
Simulated Target 10%
Raster
CommonRare
0
Noise σ

Authors

  • The University of Tokyo
  • UTIAS
  • WPI
  • IRCN
  • JSPS
  • KTH Royal Institute of Technology