Read on Twitter
#ICLR2020 is here and I am excited about our paper at #BAICS2020 so, I wrote an introduction to systematic generalization (the thing our work explores) and discussed 4 papers from ICLR 2020 on sys. gen. (including ours) in this post.
TL;DR this thread https://medium.com/@prakashkagitha/iclr-2020-papers-on-compositional-systematic-generalization-including-ours-54a45badfcae?sk=0ab411438c50e79ef59c73a7774cfee9
TL;DR this thread https://medium.com/@prakashkagitha/iclr-2020-papers-on-compositional-systematic-generalization-including-ours-54a45badfcae?sk=0ab411438c50e79ef59c73a7774cfee9
Environmental drivers of systematicity and generalization in a situated agent
Shows that more objects, the frame of reference, among others, are drivers for sys. gen.
by @FelixHill84 @AndrewLampinen Rosalia Schneider, Stephen Clark, Matt Botvinick, James McClelland, @santoroAI
Shows that more objects, the frame of reference, among others, are drivers for sys. gen.
by @FelixHill84 @AndrewLampinen Rosalia Schneider, Stephen Clark, Matt Botvinick, James McClelland, @santoroAI
See the paper, video, and join the conversation here
https://iclr.cc/virtual/poster_SklGryBtwr.html
https://iclr.cc/virtual/poster_SklGryBtwr.html
Measuring Compositional Generalization: A Comprehensive Method on Realistic Data
Their method automatically creates train/test splits with higher atom convergence and lower compound divergence to test compositional generalization
by @keysers @sinopalnikov @marcvanzee @obousquet
Their method automatically creates train/test splits with higher atom convergence and lower compound divergence to test compositional generalization
by @keysers @sinopalnikov @marcvanzee @obousquet
See the paper, video, and join the conversation here
https://iclr.cc/virtual/poster_SygcCnNKwr.html
https://iclr.cc/virtual/poster_SygcCnNKwr.html
Permutation Equivariant Models for Compositional Generalization in Language
It hypothesizes that language compositionality is a form of group-equivariance and shows good evidence for it.
by Jonathan Gordon, David Lopez-Paz, Marco Baroni, Diane Bouchacourt
It hypothesizes that language compositionality is a form of group-equivariance and shows good evidence for it.
by Jonathan Gordon, David Lopez-Paz, Marco Baroni, Diane Bouchacourt
See the paper, video, and join the conversation here
https://iclr.cc/virtual/poster_SylVNerFvr.html
https://iclr.cc/virtual/poster_SylVNerFvr.html
Systematic generalization emerges in seq2seq models with variability in data (our paper)
We show that the standard LSTM+attn model can exhibit sys.gen. and analyze it's behavior when it does.
See the paper, video, and join the conversation here
https://baicsworkshop.github.io/program/baics_11.html
We show that the standard LSTM+attn model can exhibit sys.gen. and analyze it's behavior when it does.
See the paper, video, and join the conversation here
https://baicsworkshop.github.io/program/baics_11.html
I made a summary of our work in this thread here. https://twitter.com/prakashkagitha/status/1249760207325908992?s=20
Finally, I wonder what is your opinion in enabling systematic generalization in deep learning models.
RT for more samples!
RT for more samples!
