Sylvain

ML Research @ Google Brain Zurich

[Mario Lucic @MarioLucic_]

I'm happy to announce that we have open-sourced our suite of metrics for out-of-distribution generalization and uncertainty quantification. It is framework independent and we provided TF, PyTorch and JAX examples. https://t.co/VKxoIaVlfK

[Olivier Bachem @OlivierBachem]

We spent the last months at @GoogleAI Zurich & Paris trying to understand on-policy RL for locomotion using a large scale study (>250'000 models). Check out https://t.co/acc37lMixA for insights and practical recommendations on >50 high- and low-level implementation decisions! 1/2

[Sylvain @sylvain_gelly]

Great results from our team Google Brain Zurich! Simple and scalable recipes seem to still be effective.

[Lucas Beyer (bl16) @giffmana]

We just updated our Big Transfer (BiT) paper with ObjectNet results. Pre-training of large models shows huge improvement on this "in the wild" vision task: almost +30% absolute! (https://t.co/ayVTtLobB7)

[Michael Tschannen @mtschannen]

New #CVPR2020 paper: Self-Supervised Learning of Video-Induced Visual Invariances (VIVI)

https://t.co/36X5g6J3TP

We leverage invariances at the

1) frame level,
2) shot/clip level,
3) video level,

to learn transferable image representations from raw videos in the wild. 1/4

[Olivier Bachem @OlivierBachem]

Checkout our latest work where we provide a framework to automatically remove shortcuts for self-supervised learning.

[Sylvain @sylvain_gelly]

Will 2020 see more pretrained representation and transfer works? This simple (but large scale) BiT approach is quite effective on a wide number of datasets. You can see here *all* top-1 errors on CIFAR10 test. VTAB testbed is improved but still challenging https://t.co/xweZb6dKk3

[Lucas Beyer (bl16) @giffmana]

We distill key components for pre-training representations at scale: BigTransfer ("BiT") achieves SOTA on many benchmarks with ResNet, e.g. 87.8% top-1 on ImageNet (86.4% with only 25 images/class) and 99.3% on CIFAR-10 (97.6% with only 10 images/class). https://t.co/2iirPo1G45

[Sjoerd van Steenkiste @vansteenkiste_s]

An update of our paper investigating object compositionality in GANs is now available: https://t.co/4pfREqh17O

We show how a structured generator that learns about objects can facilitate unsupervised instance segmentation.

w/ @karol_kurach @SchmidhuberAI @sylvain_gelly

1/4

[Samaneh Azadi @smnh_azadi]

Using our Semantic Bottleneck GAN (SB-GAN), achieve SOTA results in synthesizing complex scenes from scratch as well as from real semantic layouts.
Joint work with @mtschannen,Eric Tzeng,@sylvain_gelly, Trevor Darrell, @MarioLucic_
https://t.co/afiuCPKnx9
https://t.co/1Y1AMtVjfo

[Mario Lucic @MarioLucic_]

In our recent collaboration with @berkeley_ai we show how to generate realistic complex scenes from scratch! While the problem is extremely challenging, we show how to achieve SOTA in unconditional generation and improve conditional generation using SPADE https://t.co/YFdyuLksce

[Google AI @GoogleAI]

We are happy to announce the v2.0 release of the Google Research Football Environment. The most exciting feature of this release is the Game Server, which lets your agent compete online with other researchers' models. Visit https://t.co/2ow3q98ZOP and give it a try!

[Google AI @GoogleAI]

We’re pleased to release the Visual Task Adaptation Benchmark (VTAB), a diverse, realistic, and challenging protocol to measure progress towards universal visual representations. Learn all about it below. https://t.co/PbORwSFPAg

[Sylvain @sylvain_gelly]

Are you interested in Representation Learning, Transfer Learning, Domain Adaptation, Self-Supervised Learning or Semi-Supervised Learning? Have a look at this work from Google Brain Zurich!

[Neil Houlsby @neilhoulsby]

How effective is representation learning? To help answer this, we are pleased to release the Visual Task Adaptation Benchmark: our protocol for benchmarking any visual representations (https://t.co/IQkVRha8vh) + lots of findings in https://t.co/Ai8uKW9AlS.
@GoogleAI

[Mario Lucic @MarioLucic_]

Excited about recent progress in self-supervised representation learning based on mutual information maximisation? Mutual Information might not be the key ingredient for the success of these methods, as shown in our latest paper: https://t.co/Ujy7g4I3mK

[Neil Houlsby @neilhoulsby]

Code released to adapt BERT using few parameters. Can be used to adapt one model to many tasks. Catastrophic forgetting not included. https://t.co/hrRIVHpQno

[Google AI @GoogleAI]

Congratulations to the Google, @ETH and @MPI_IS authors of "Challenging Common Assumptions in the Unsupervised Learning of Disentangled Representations" (https://t.co/72TzrG0rvV), recipient of an #ICML2019 Best Paper Award! Learn more in the blog post at https://t.co/lrC7eN4EUx.

[Google AI @GoogleAI]

Check out a novel #ReinforcementLearning environment where agents aim to master the world’s most popular sport—football! The Google Research Football Environment includes benchmarks & progressive RL training scenarios, and is available in open source beta→https://t.co/rdvsOINA9e

[Mario Lucic @MarioLucic_]

Quantitative evaluation of generative models is a key research challenge. In our latest work, we introduce a theoretical framework which uncovers the existing approaches based on precision and recall as special cases, and offers novel geometric insights. https://t.co/2AjXNG9DeT

[Xiaohua Zhai @XiaohuaZhai]

Want to turn your self-supervised method into a semi-supervised learning technique? Check out our S⁴L framework (https://t.co/D14Og5SDxK)!

Work done at @GoogleAI with @avitaloliver, @__kolesnikov__ and @giffmana.

[Mario Lucic @MarioLucic_]

Our work on "High-Fidelity Image Generation With Fewer Labels" has been accepted to ICML'19! Thanks to the reviewers and the area chairs for the thorough reviews! @icmlconf @mtschannen @TheMarvinRitter @XiaohuaZhai @OlivierBachem @sylvain_gelly @GoogleAI

[Google AI @GoogleAI]

Learning disentangled representations of a scene is critical for many machine vision tasks. In collaboration with @ETH and @MPI_IS, we present a broad examination of the field, examine the role of implicit biases and provide direction for future research. https://t.co/0VnIhLFJYj