Working on Greater Generalization in Artificial Intelligence
The conference season begins today with the Tenth International Conference on Learning Representations (International Foreign Language Association 2022), which will be held virtually April 25–29, 2022. Participants from around the world come together to share their cutting-edge work in representational learning, from developing cutting-edge AI solutions to data science, machine vision, robotics, and more.
On the first day of the conference, Pushmeet Kohli, our AI for Science and Tough and Verified AI lead, will deliver a talk on how AI can radically improve solutions to a wide range of scientific problems, from genomics and structural biology to quantum chemistry and even pure mathematics.
In addition to supporting the event as sponsors and regular workshop organizers, our research teams are presenting 29 papers, including 10 collaborations this year. Here is a quick overview of our upcoming oral, spotlight, and poster presentations:
Optimizing learning
Many key papers focus on key ways we are improving the learning process of our AI systems. This includes increasing efficiency, developing “fewer shots” learning, and creating proficient data systems that reduce computational costs.
IN “Bootstrapped Meta-Learning”some ICLR 2022 Outstanding Paper Award winner, we propose an algorithm that enables the agent to learn how to learn by learning itself. We also present policy improvement algorithm which redesigns AlphaZero – our self-taught system for chess, shogi, and go – to continue to improve, even when trained with a diminutive number of simulations; a regularizer that mitigates the risk of bandwidth loss in a wide range of RL agents and environments; and improved architecture enabling efficient training of attention models.
Test
Curiosity is a key part of human learning, helping to develop knowledge and skills. Similarly, exploration mechanisms allow AI agents to go beyond existing knowledge and discover the unknown or try something modern.
Moving on to the question “When should agents start exploring?”, we investigate when agents should switch to exploration mode, at what time scales it makes sense to switch, and which signals best determine how long and frequent exploration periods should be. In another paper, we introduce “bonus for information exploration“allowing agents to break through the constraints of real-life intrinsic rewards and learn more skills.
Robust AI
To deploy ML models in the real world, they must be effective when switching between training, testing, and new data sets. Understanding the causal mechanisms is essential, allowing some systems to adapt while others struggle to cope with new challenges.
Extending the research on these mechanisms, we present an experimental framework that allows for a detailed analysis analysis of resistance to distribution changes. Robustness also helps protect against adversarial damage, whether it is unintentional or intentional. In the case of image damage, we propose a technique that theoretically optimizes parameters of image-to-image models to reduce blur, fog and other common problems.
Emerging Communication
AI agents not only help machine learning researchers understand how agents develop their own communication to complete tasks, but they can also discover information about linguistic behaviors within populations, which could lead to more interactive and useful AI.
Working with researchers from Inria, Google Research, and Meta AI, we connect the dots between the role of diversity in human populations in shaping language partially resolve the apparent contradiction in computer simulations with neural agents. Then, because building better representations of language in AI is so important to understanding emerging communication, we also study the importance of scaling dataset, task complexity and population size as independent aspects. Furthermore, we also investigated compromises between expression, complexity and unpredictability in games where multiple agents communicate with each other to achieve a single goal.
See the full scope of our work at ICLR 2022 Here.