Thomas Kipf: Excited to share our work on Contrastive Learning of Structured World Models!
C-SWMs learn object-factorized models & discover objects without supervision, using a simple loss inspired by work on graph embeddings
3 replies, 542 likes
Thomas Kipf: Our paper on Structured World Models got accepted to #ICLR2020 as a long talk! @iclr_conf
Congrats to amazing co-authors @ElisevanderPol @wellingmax
5 replies, 283 likes
hardmaru: Contrastive Learning of Structured World Models
A structured understanding of our world in terms of objects, relations and hierarchies is an important part of human cognition.
This paper explores using graph neural nets to learn structured world models.
1 replies, 240 likes
mat kelcey: been hacking for awhile on the idea of training an embedding based YOLO with self supervision contrastive loss. have never had a good solution for unsupervised object detection piece.... until now!!!
"Contrastive Learning of Structured World Models" https://arxiv.org/abs/1911.12247 https://t.co/k2h4I6mkrF
1 replies, 167 likes
Elise van der Pol: 👇This! Happy with our new work on learning object-factorized models without labels.
0 replies, 43 likes
Adam Kosiorek: very cool paper :)
0 replies, 27 likes
Jesse Engel: I'm a big fan of object oriented generative/world models so it's great to see all the progress in this area.
0 replies, 25 likes
Sjoerd van Steenkiste: Interesting to see that one can learn about objects without reconstructing in pixel-space! I wonder if contrastive learning in this way will allow us to overcome the limitations of existing object-centric approaches in coping with complex backgrounds that are hard to reconstruct.
0 replies, 17 likes
Daniel Daza: Unsupervised learning: ✅
Object decomposition: ✅
No ELBO or Markov chains: ✅
Contrastive loss in latent space: ✅
Plenty of reasons to be excited!
0 replies, 10 likes
hardmaru: @ngutten @yudapearl @jsusskin @LeCun I think @thomaskipf’s paper that looks at building unsupervised, structured world models might be also a good way forward at evaluating unseen interventions. In their paper they propose a metric that might have similar motivations as this discussion~ https://twitter.com/thomaskipf/status/1200000165068902400?s=21
1 replies, 9 likes
hardmaru: @ngutten @yudapearl @jsusskin @LeCun I think @thomaskipf’s paper that looks at building unsupervised, structured world models might be also a good way forward at evaluating unseen interventions. In their paper they propose a metric that might have similar motivations as this discussion. https://arxiv.org/abs/1911.12247
0 replies, 6 likes
hardmaru: @yujin_tang @karpathy Here's a cool work using graph neural nets to model relationships of discrete entities extracted from pixels as part of the learning process. They also generalize better to novel scenes.
But while entities are discrete, embeddings are still real valued... https://twitter.com/thomaskipf/status/1200000165068902400
0 replies, 5 likes
Found on Nov 28 2019 at https://arxiv.org/pdf/1911.12247.pdf