Best paper award at #ICML2018 main idea: identify a class of adversarial defenses, called "Obfuscated gradients", & develop attack techniques to thwart this defense. Circumvents 6 of 9 defenses presented at #ICLR2018.
— Reza Zadeh (@Reza_Zadeh) July 8, 2018
Paper: https://t.co/Z40rE2qKIx
Code: https://t.co/uDZxycCDN7
Interesting! Adding that to the reading list. Is there a particular reason they didn't include the recent GroupNorm by Yuxin Wu & Kaiming He? https://t.co/UMSiG3kxXU
— Sebastian Raschka (@rasbt) July 8, 2018
The quest for optimal normalization in neural nets continues. SwitchNorm: add BatchNorm + InstanceNorm + GroupNorm with a learnable blend at each layer https://t.co/9hOQXnkk8T fun plots; + code https://t.co/34r96BStCS
— Andrej Karpathy (@karpathy) July 7, 2018
How to Backdoor Federated Learning, Bagdasaryan et al. – attacks federated learning scenarios where many users contribute to a single shared model: https://t.co/7liNzrenBa pic.twitter.com/c8f7xEds04
— Brendan Dolan-Gavitt (@moyix) July 7, 2018
Will present my #midl2018 poster next session. Learn how to train a normal CNN with 8192x8192 input sizes and a single label on one GPU! (from 235gb to 7gb memory required)
— Hans Pinckaers (@hanspinckaers) July 6, 2018
Code: https://t.co/KzDPFnbArl pic.twitter.com/Vcfco5fBPk
Few-shot learning of neural networks from scratch by pseudo example optimization. (arXiv:1802.03039v3 [https://t.co/zjV5HgYw5a] UPDATED) https://t.co/MGuQpVJQnY
— Stat.ML Papers (@StatMLPapers) July 6, 2018
Given a satellite image, machine learning creates the view on the ground - Interesting work https://t.co/2Cypnw6o4w
— Nando de Freitas (@NandoDF) July 5, 2018
SLIM can model a scene conditioned only on human-written language descriptions. This allows it to capture the semantic meaning of spatial relationships in language. @tmramalho @TomasKocisky @arkitus @karlmoritz https://t.co/nAvFBKcb61 pic.twitter.com/sikxhXaLCg
— DeepMind (@DeepMindAI) July 5, 2018
We have significantly updated our preprint on Variance Networks! Training only variances of the weights w_ij ~ N(0, sigma_ij^2) leads to state-of-the-art performance. Now more explanation, practical and theoretical results. https://t.co/5KobBhmHLA https://t.co/WTAnJoX4H6
— Arsenii Ashukha (@arsashuha) July 5, 2018
This new family of GAN loss functions looks promising! I'm especially excited about Fig 4-6, where we see that the new loss results in much faster learning during the first several iterations of training. I implemented the RSGAN loss on a toy problem and it worked well. https://t.co/BgVPQZE4Nx
— Ian Goodfellow (@goodfellow_ian) July 3, 2018
How can one detect friendly and adversarial behaviour from raw data? https://t.co/cRJCyfo5oo
— DeepMind (@DeepMindAI) July 3, 2018
"Co-Training of Audio and Video Representations from Self-Supervised Temporal Synchronization," Korbar et al.: https://t.co/XlOizuXwn2
— Miles Brundage (@Miles_Brundage) July 3, 2018
Another clever idea for video self-supervision, like the colorization one from a few days ago
