Next week, the 40th International Conference on Machine Learning (ICML 2023) will begin in Honolulu, Hawai’i. This conference brings together the AI community to share new ideas, tools, and datasets, and to make connections that will advance the field. Researchers from around the world will be presenting their latest advancements in areas such as computer vision and robotics.
One of our directors, Shakir Mohamed, will be giving a talk at ICML on machine learning with a social purpose. He will be addressing challenges in healthcare and climate by taking a sociotechnical view and strengthening global communities.
We are proud to be a Platinum Sponsor of the conference and to continue our collaboration with long-term partners such as LatinX in AI, Queer in AI, and Women in Machine Learning.
During the conference, we will be showcasing demos of some of our latest projects. This includes AlphaFold, which is our AI system for protein structure prediction. We will also be presenting our advancements in fusion science and introducing new models like PaLM-E for robotics and Phenaki for generating videos from text.
This year, Google DeepMind researchers will be presenting over 80 new papers at ICML. As some of these papers were submitted before Google Brain and DeepMind joined forces, we will be featuring the papers submitted under a Google Brain affiliation on the Google Research blog, while this blog will feature papers submitted under a DeepMind affiliation.
Now let’s take a closer look at some of the exciting topics that will be discussed at ICML:
AI in the (simulated) world
Our latest research focuses on translating the capabilities of AI systems from simulated environments into the real world. We are developing AI agents, such as AdA, that can adapt to solve new problems in simulated environments, similar to how humans do. These agents can take on challenging tasks like combining objects in novel ways, navigating unseen terrains, and cooperating with other players. Additionally, we are exploring the use of vision-language models to train embodied agents, for example, by providing instructions to robots.
The future of reinforcement learning
To ensure responsible and trustworthy AI, it is crucial to understand the goals behind these systems. In reinforcement learning, we can define goals through reward. At ICML, we aim to settle the reward hypothesis first proposed by Richard Sutton, which states that all goals can be thought of as maximizing expected cumulative reward. We will explore the conditions under which this holds true and clarify the objectives that can be captured by reward in the context of reinforcement learning. We will also discuss how to train reinforcement learning algorithms within constraints to make them robust for real-world deployment. Additionally, we will demonstrate how models can learn complex long-term strategies in games with imperfect information, like poker.
Challenges at the frontier of AI
Humans possess the ability to learn, adapt, and understand the world around us. Developing advanced AI systems that can generalize in human-like ways will unlock new possibilities for AI tools that can be used in everyday life and tackle new challenges. Our research focuses on plasticity in neural networks and how it can be preserved over the course of training. We also study meta-trained neural networks to understand the in-context learning that emerges in large language models. In addition, we introduce a new family of recurrent neural networks that performs better on long-term reasoning tasks. Lastly, we propose an approach called “quantile credit assignment” to help AI systems better understand complex, real-world environments by disentangling luck from skill.
In conclusion, ICML 2023 is an exciting event that brings together researchers and experts in AI from around the world. It provides a platform for sharing new ideas and advancements in the field. We are proud to be a part of this conference and to contribute to the ever-evolving field of AI.