NeurIPS 2022: A Showcase of Cutting-Edge AI Research
Advancing best-in-class large models, compute-optimal RL agents, and more transparent, ethical, and fair AI systems
The thirty-sixth International Conference on Neural Information Processing Systems (NeurIPS 2022) is set to take place from 28th November to 9th December 2022 in New Orleans, USA, as a hybrid event welcoming AI and ML enthusiasts from all around the world.
As Diamond sponsors of NeurIPS 2022, DeepMind is honored to support the event and contribute significantly to the exchange of research advances in the AI and ML community. With a commitment to innovation and ethical AI development, teams from DeepMind are presenting 47 papers, including 35 external collaborations in virtual panel sessions and poster presentations.
Breaking Ground with Large Models
Large models (LMs) are transforming the AI landscape with their exceptional performance in various domains such as language, text, audio, and image generation. DeepMind’s Chinchilla stands out as a 70 billion parameter language model that surpasses larger models like Gopher by leveraging updated scaling laws for more effective training. This pioneering work has already influenced the development of leaner and more efficient models, earning DeepMind an Outstanding Main Track Paper award at the conference.
In addition to Chinchilla, the team is introducing Flamingo, a family of few-shot learning visual language models that excel in handling diverse data types. This innovative approach bridges the gap between vision-only and language-only models, achieving state-of-the-art results in few-shot learning tasks across multiple modalities.
Optimizing Reinforcement Learning
DeepMind continues to push the boundaries of reinforcement learning (RL) with novel approaches to enhance the decision-making capabilities of RL agents. By expanding the scale of information accessible for retrieval, these agents deliver improved performance in a computationally efficient manner.
Among the highlights is BYOL-Explore, an RL agent designed for curiosity-driven exploration in visually complex environments. This agent achieves superhuman performance while maintaining robustness and simplicity, setting a new standard in RL research.
Advances in Algorithms
DeepMind’s commitment to algorithmic innovation is evident in its groundbreaking work on automating the configuration of computer networks using neural algorithmic reasoning. This flexible approach demonstrates a remarkable speed improvement compared to existing methods, emphasizing the importance of incremental algorithmic improvements in saving energy and resources.
Moreover, the team explores the concept of “algorithmic alignment” and its implications for optimizing out-of-distribution performance. Through a rigorous examination of graph neural networks and dynamic programming, DeepMind sheds light on the intricate relationship between these components.
Pioneering Responsible AI
As responsible pioneers in the field of AI, DeepMind prioritizes transparency, ethics, and fairness in AI development. Understanding and explaining the behavior of complex AI systems is essential for creating fair and accurate models, paving the way for safer and more ethical AI applications.
DeepMind’s pursuit of fair AI extends to developing AI systems capable of self-explanation, fostering a deeper understanding of their decision-making processes. By introducing concepts like counterfactual harm and exploring methods to diagnose and mitigate fairness failures, DeepMind is at the forefront of promoting ethical AI deployment, particularly in critical domains like healthcare.
Experience the full spectrum of innovative research from DeepMind at NeurIPS 2022 here.