DeepMind Presents at ICLR 2022 Conference
DeepMind’s Head of AI for Science and Robust and Verified AI teams, Pushmeet Kohli, is speaking at ICLR 2022, where experts are sharing advancements in AI and representational learning. DeepMind is also presenting 29 papers, focusing on various aspects of AI.
Optimizing Learning
DeepMind’s papers aim to improve the learning process of AI systems by enhancing performance, few shot learning, and reducing computational costs. The “Bootstrapped meta-learning” paper, an ICLR 2022 Outstanding Paper award winner, introduces an algorithm that enables an agent to teach itself how to learn. Other papers focus on policy improvement, regularizers, and improved architectures for training attentional models.
Exploration
Curiosity drives human learning, and exploration mechanisms allow AI agents to go beyond existing knowledge. DeepMind’s research investigates when agents should explore and introduces an “information gain exploration bonus” to enhance learning of new skills.
Robust AI
Effective deployment of ML models in the real world requires robustness when shifting between training, testing, and new datasets. DeepMind’s research focuses on understanding causal mechanisms and protecting against adversarial harms such as image corruptions.
Emergent Communication
AI agents have the potential to reveal insights into linguistic behaviors within populations. DeepMind’s research explores the role of diversity within human populations on shaping language and investigates the importance of scaling up the dataset, task complexity, and population size as independent aspects in AI’s language representations.
For more information, visit the ICLR 2022 website.