The International Conference on Neural Information Processing Systems (NeurIPS 2022) is happening from 28 November – 9 December 2022 in New Orleans, USA. It’s the world’s biggest conference in artificial intelligence and machine learning, and DeepMind is proud to be Diamond sponsors.
At the event, DeepMind is presenting 47 papers, including collaborations with external partners. They’ve made significant advancements in large models, reinforcement learning, algorithmic advances, and responsible AI.
Advancements in large models: DeepMind has created a 70 billion parameter language model that outperforms many larger models. They’ve also introduced Flamingo, a family of few-shot learning visual language models that sets a new state of the art in few-shot learning on a wide range of multimodal tasks.
Optimizing reinforcement learning: DeepMind has introduced a new approach that boosts the decision-making abilities of RL agents in a compute-efficient way. They’ve also developed an RL agent called BYOL-Explore, which achieves superhuman performance while being robust to noise and simpler than prior work.
Algorithmic advances: DeepMind has shared a highly scalable method for the automatic configuration of computer networks and explored the relationship between graph neural networks and dynamic programming for optimizing out-of-distribution performance.
Pioneering responsibly: DeepMind is committed to developing AI systems that are transparent, ethical, and fair. They’ve introduced a practical way to explain the behavior of complex AI systems and a statistical measure called counterfactual harm to avoid pursuing harmful policies. They’ve also proposed ways to diagnose and mitigate failures in model fairness caused by distribution shifts.
You can see the full range of DeepMind’s work at NeurIPS 2022 here.