Thomas Kipf: Excited to share our work @GoogleAI on Object-centric Learning with Slot Attention!
Slot Attention is a simple module for structure discovery and set prediction: it uses iterative attention to group perceptual inputs into a set of slots.
14 replies, 731 likes
Thomas Kipf: Happy to learn that our work on Slot Attention has been accepted for spotlight presentation at @NeurIPSConf!
4 replies, 388 likes
Thomas Kipf: We have released the code for Slot Attention (incl. pre-trained model checkpoints on CLEVR)
NeurIPS camera ready: https://arxiv.org/abs/2006.15055 https://t.co/1lNagRiKoy
1 replies, 386 likes
Francesco Locatello: Super excited to share what I’ve been working on in the past months during my internship at Google Brain in Amsterdam: "Object-Centric Learning with Slot Attention" https://arxiv.org/pdf/2006.15055.pdf @GoogleAI [1/7] https://t.co/1aNYLm4exj
4 replies, 341 likes
Thomas Kipf: SCOUTER: An explainable image classifier using a modified version of Slot Attention
by Liangzhi Li et al. (Osaka University)
Slot Attention: https://arxiv.org/abs/2006.15055 https://t.co/u6MfKTwoMj
2 replies, 206 likes
Alexandr Kalinin: There is already a #PyTorch implementation of the Slot Attention module:
0 replies, 187 likes
Francesco Locatello: We have finally released the code for Slot Attention (+ some checkpoints) and updated the paper!
#NeurIPS2020 camera ready: https://arxiv.org/abs/2006.15055 https://t.co/SyqWuJckv5
0 replies, 124 likes
Francesco Locatello: Really cool to see Slot Attention powering explainable image classifiers!
When writing the broader impact statement for #NeurIPS2020, we were hoping to see developments in this direction. I'm thrilled Liangzhi et al. did it.
1 replies, 42 likes
Andrew Davison: Slot attention sounds like an interesting concept for structure discovery.
0 replies, 24 likes
Anirudh Goyal: I liked this work by @thomaskipf and @FrancescoLocat8. They use RIM style slot based top-down attention, with the caveat that all the "slots" share the same parameters (via recurrence also, and that provides equivariance).
I'm glad you guys tried it! :)
0 replies, 15 likes
Peter Steinbach: Does this mark the ☀️rise of unsupervised segmentation? Would love to try this with scientific data opposed to natural scenes! Volunteers? @martweig @uschmidt83 @sagzehn @helmholtz_ai @noreenwalk @haesleinhuepf Still crawling this if multiple instances of a single class work too.
2 replies, 13 likes
Sungjin Ahn: So excited that we have @thomaskipf as our invited speaker in the ICML Workshop on Object-Oriented Learning (WOOL)! Looking forward to the talk! Join the ICML WOOL!
0 replies, 9 likes
Patrick Emami: Transformer-like attention can be used for perceptual grouping! We saw it used prev for learning patch-like object-part representations via capsules in SCAE.
This approach seems to lose IODINE’s translation equivariance tho—important for sequences 🤔
1 replies, 4 likes
SLAM-Hub: Object-Centric Learning with Slot Attention (Under review)
Paper : https://arxiv.org/abs/2006.15055
0 replies, 2 likes
Daisuke Okanohara: Slot attention is a map from N input feature vectors to K output vectors (slots) using dot-product attention and iterative routing (c.f., capsule), and is input-perm. invariant and output-perm. equivariant, ideal for representing a
set of objects. https://arxiv.org/abs/2006.15055
0 replies, 2 likes
Found on Jun 29 2020 at https://arxiv.org/pdf/2006.15055.pdf