Flamingo seamlessly handles input sequences of images, videos, & text by bridging powerful pretrained LMs with visual encoders. With only simple few-shot learning (no fine-tuning) a single 🦩 achieves SotA results on several competitive computer vision benchmarks in 32 shots. 2/

0:22

4.8K views

2

14

100

Introducing Flamingo 🦩: a generalist visual language model that can rapidly adapt its behaviour given just a handful of examples. Out of the box, it's also capable of rich visual dialog. Read more: dpmd.ai/dm-flamingo 1/

12

292

913

New work from our safety team proposes counterfactual reasoning as a key ingredient for safe and ethical AI. Read more: dpmd.ai/counterfactual Work by

@jonathanrichens

, Rory Beard (

@CoMind_

) and Daniel Thompson (Babylon).

arxiv.org

First do no harm: counterfactual objective functions for safe...

To act safely and ethically in the real world, agents must be able to reason about harm and avoid harmful actions. In this paper we develop the first statistical definition of harm and a framework...

5

26

94

Topics to follow

Sign up to get Tweets about the Topics you follow in your Home timeline.

Carousel

Work by

, Oleg Sushkov, Jean-Baptiste Regli, Stefan Schaal,

@yusufaytar

,

@markus_with_k

and Jon Scholz. #ICLR2022

2

20

Work by

, Adrià Puigdomenech, Jacob Walker, Tim Scholtes,

@jovana_mitr

&

@BlundellCharles

2

11

CoBERL, a simple and scalable method to improve data efficiency in a variety of RL environments (Atari, DmControl, DmLab). Lean more today dpmd.ai/ICLR-CoBERL #ICLR2022

1

26

77

Today

, see how an RL agent can use demonstration based goal conditioning to effectively learn to perform complex long-horizon manipulation tasks. dpmd.ai/ICLR-Hindsight #ICLR2022

3

19

74

Work by

,

,

,

, András György,

,

2

10

New work that adversarially optimises pre-trained autoencoders to corrupt inputs and improve robustness to common corruptions. Join the poster session today at 10:30AM BST: dpmd.ai/ICLR-Adversari #ICLR2022

1

15

49

This method enables users to nearly match the speed and *generalisation* performance of ResNets using vanilla convents. Work by

@Guodzh

, Aleksandar Botev & James Martens #ICLR2022

2

22

Research shows a dramatic improvement in the trainability of difficult-to-train architectures by “tailoring” rectifier activations for them: dpmd.ai/ICLR-Tailored- #ICLR2022

1

18

92

Instead of L2R prompting, SUNDAE can fill in arbitrary patterns in a template or update a complete text draft - hinting at new capabilities for creative text co-editing between humans & machines! By

,

,

,

&

4

22

With SUNDAE, our team shows how unrolling the autoencoder denoising process during training allows users to achieve strong results in machine translation & language modelling. Join the session at 6:30PM BST: dpmd.ai/SUNDAE #ICLR2022

GIF

2

14

68

Today

learn more about a mechanism by which non-stationary prediction targets can prevent learning progress in deep RL agents. By Clare Lyle, Mark Rowland, and

@wwdabney

. See the schedule for more: dpmd.ai/ICLR-2022 #ICLR2022

deepmind.events

ICLR 2022 | Gold Sponsor

deepmind events

12

43

Work by

, David Szepesvari, Georg Ostrovski, Diana Borsa, and Tom Schaul. #ICLR2022 2/2

1

12

This morning

- an adaptive mechanism which endows an agent with diverse and useful experiences that translate to meaningful performance differences. Learn more at 10:30AM BST: dpmd.ai/ICLR-Explorati #ICLR2022 1/

1

7

49

Work by

,

,

,

,

, and

. #ICLR2022 2/2

1

14

Can model-based RL generalise better than model-free approaches in practice? Yes! Research finds that MuZero with self-supervision is SotA on Procgen with 10x less data, implicit meta-RL on MetaWorld. Learn more at the session tonight:dpmd.ai/ICLR-Procedura #ICLR2022 1/

2

23

116

Work by

, Yannick Schroecker,

@TZahavy

,

@hado

, David Silver, and Satinder Singh. #ICLR2022 #OutstandingPaperAward 2/2

1

8

How can an agent learn *to learn* from its own experiences? Today

@iclr_conf

our team shows how an agent can use its own future learning as a teacher with Bootstrapped Meta-Learning. Oral session at 18:30PM BST: dpmd.ai/ICLR-Bootstrap #ICLR2022 #OutstandingPaperAward 1/

Quote Tweet

DeepMind

@DeepMind

· Apr 22

Congratulations to the team for their #ICLR2022 Outstanding Paper Award. Opening a new direction in meta-learning, “Bootstrapped meta-learning”, proposes an algorithm that enables an agent to learn how to learn by teaching itself: dpmd.ai/Outstanding-Pa 1/

Show this thread

2

18

66

Ever wonder why foundation models like ImageNet classifiers are useful in new tasks? Our team explains this phenomenon through the notion of neural collapse today at 6:30PM BST: dpmd.ai/ICLR-Neural-Co Work by

@GalantiTomer

, András György, and

@mhutter42

#ICLR2022

13

48

DeepMind

@DeepMind

Apr 26

Work by James Martens,

,

,

, Valentin Dalibard,

(Google Brain), and

(Google Brain) Code: dpmd.ai/DKS-Code 2/2

github.com

GitHub - deepmind/dks: Multi-framework implementation of Deep Kernel Shaping and Tailored Activat...

Multi-framework implementation of Deep Kernel Shaping and Tailored Activation Transformations, which are methods that modify neural network models (and their initializations) to make them easier to...

6

38

By carefully transforming neural networks they can be trained much more easily. When combined with K-FAC, this method allows vanilla convnets (w/o shortcut connections or normalization layers) to be trained as fast as standard ResNets of the same depth! dpmd.ai/DKS-Paper 1/

3

49

276

Work by Jonathan Godwin, @m_shcaarschmidt, Alexander Gaunt, Alvaro Sanchez-Gonzalez, Yulia Rubanova,

@PetarV_93

, James Kirkpatrick, and

@PeterWBattaglia

. #ICLR2022

2

16

The Noisy Nodes method is a simple oversmoothing regularisation technique that significantly improves results for MPNNs for 3D and 2D molecular property prediction. Learn more in today’s session: dpmd.ai/ICLR-Noisy-Nod #ICLR2022

2

12

73

Work by Shantanu Thakoor,

,

,

,

, Remi Munos,

and

. #ICLR2022

1

21

Bootstrapped Graph Latents - a self-supervised representation learning method that does not require negative examples and naturally scales to extremely large graphs. Join the poster session today: dpmd.ai/ICLR-Graph-Lat #ICLR2022

1

23

114

Work by

,

,

and

. #ICLR2022

1

2

13

Today

: a MCTS inspired algorithm that can find high-scoring outputs under translation models that make no assumptions about the form or structure of the search objective. Learn more: dpmd.ai/ICLR-ATS #ICLR2022

1

16

59

Work by Siqi Liu,

, Daniel Hennes, Josh Merel, Nicolas Heess and

@ThoreG

. Demo: dpmd.ai/NeuPL-Demo #ICLR2022

neupl.github.io

NeuPL: Neural Population Learning

Siqi Liu1,2, Luke Marris1,2, Daniel Hennes2, Josh Merel2, Nicolas Heess2, Thore Graepel1,2

3

22

NeuPL: a method that preserves game-theoretic properties of population learning and yields a single conditional network that learns and represents a population of diverse policies. Why is that important? Learn more at 10:30AM BST! dpmd.ai/ICLR-NeuPL #ICLR2022

3

9

94

Work by

, Florian Strub, Florent Altché,

, Eugene Trasso,

,

, @olivertieleman,

and Bilal Piot #ICLR2022

3

14

Later today

- find out how important scaling up the emergent communication framework is to the emergence of more productive and robust languages. 👉👉 dpmd.ai/ICLR-Emergent- #ICLR2022

2

4

34

FNF: a new fair representation learning method with fairness guarantees via distribution alignment. Presented today at

by

,

and

(with

). Paper: dpmd.ai/fnf Code: dpmd.ai/fnf-code #ICLR2022

1

25

97

By

, Brian McWilliams,

&

#ICLR2022 2/2

1

12

Tired of playing with loaded dice? In EigenGame Unloaded, our team shows how agents playing a game can scale PCA to datasets with billions of samples and large social networks. Join the poster session today at 6:30PM BST: dpmd.ai/ICLR-EigenGame #ICLR2022 1/

1

21

118

Show this thread

Mar	APR	May
	29
2021	2022	2023

DeepMind’s Tweets

Topics to follow

Carousel