Exploring the safety, adaptability and effectiveness of artificial intelligence for the real world
Next week marks the start of the 40th International Conference on Machine Learning (ICML 2023), to be held July 23-29 in Honolulu, Hawaii.
ICML brings together the artificial intelligence (AI) community to share new ideas, tools, and datasets and create connections to advance the field. From computer vision to robotics, researchers from around the world will present their latest advances.
Our director of science, technology and society, Shakir Mohamed, will give one talk about machine learning with a social purposeaddressing healthcare and climate challenges by taking a socio-technical perspective and strengthening global communities.
We are proud to support the conference as a Platinum Sponsor and continue to work together with our long-standing partners LatinX to AI, Queer in AIand Women in Machine Learning.
At the conference, we also present demonstrations AlphaFoldour advances in fusion science and new models such as PalM-E for robotics and Fainaki to create video from text.
Google DeepMind researchers are presenting more than 80 new papers at ICML this year. Since many papers were submitted before Google Brain and DeepMind joined forces, papers originally submitted as part of a Google Brain collaboration will be included in Google Research Blogwhile this blog includes work submitted as part of a DeepMind collaboration.
AI in the (simulated) world
The success of AI that can read, write and create is underpinned by core models – AI systems trained on massive data sets that can learn to multitask. Our latest research explores how we can translate these efforts into the real world and lays the groundwork for more generally competent and embedded AI agents that can better understand the dynamics of the world, opening up new possibilities for more useful AI tools.
In an oral presentation, we present AdA, an artificial intelligence agent that can adapt to solve new problems in a simulated environment, just like humans do. Within minutes, AdA can take on challenging tasks: combining objects in new ways, navigating unseen terrain, and collaborating with other players
Similarly, we show how we could use vision language models to help train embedded agents; – for example, telling a robot what to do.
The future of reinforcement learning
To develop responsible and trustworthy artificial intelligence, we need to understand the goals at the heart of these systems. In reinforcement learning, one way this can be defined is through reward.
In an oral presentation, we aim to settlement of the reward case first posed by Richard Sutton stating that all objectives can be thought of as maximizing expected cumulative reward. We explain the precise conditions under which it is valid and specify the kinds of goals that can – and cannot – be obtained by rewarding a general form of the reinforcement learning problem.
When developing AI systems, they need to be robust enough for the real world. We’re looking at how to do it better train reinforcement learning algorithms within constraintsas AI tools often need to be limited for safety and effectiveness.
In our research, which was identified with a ICML 2023 Outstanding Paper Awardwe investigate how we can teach models of complex long-term strategy under uncertainty with imperfect information games. We share how models can play to win two-player games even without knowing the position and possible moves of the other player.
Challenges at the Frontiers of AI
Humans can easily learn, adapt and understand the world around us. Developing advanced AI systems that can generalize in human-like ways will help create AI tools that we can use in our daily lives and tackle new challenges.
One way AI adapts is by quickly changing its predictions in response to new information. In an oral presentation, we examine plasticity in neural networks and how it can be lost during training – and ways to prevent loss.
We also present research that could help explain the type of in-context learning that occurs in large language models by studying neural networks post-trained on data sources whose statistics change spontaneously, as in natural language prediction.
In an oral presentation, we present a new family of recurrent neural networks (RNN) that perform better in long-term reasoning tasks to unlock the promise of these models for the future.
Finally, in ‘assignment of creditWe propose an approach to distinguish luck from skill. By establishing a clearer relationship between actions, outcomes, and external factors, AI can better understand complex real-world environments.