Improving best-in-class gigantic models, computationally optimal RL agents, and more lucid, ethical, and fair AI systems
Thirty-sixth International Conference on Neural Information Processing Systems (NeurIPS 2022) will take place from November 28 to December 9, 2022 as a hybrid event based in Fresh Orleans, USA.
NeurIPS is the world’s largest conference dedicated to artificial intelligence (AI) and machine learning (ML). We are proud to support this event as Diamond sponsors, helping to foster the exchange of research advances within the AI and ML community.
Teams from across DeepMind present 47 papers, including 35 projects developed in external collaborations, in virtual panels and poster sessions. Here is a brief introduction to some of the research we are presenting:
Best in class gigantic models
Vast models (LM) – generative artificial intelligence systems trained on massive amounts of data – have resulted in incredible performance in areas such as language, text, audio and image generation. Part of their success comes from sheer scale.
However, at Chinchilla we created A 70 billion parameter language model that outperforms many larger modelsincluding Gopher. We updated the scaling laws for gigantic models, showing how previously trained models were too gigantic for the amount of training performed. This work has already shaped other models that follow these updated principles, creating leaner and better models, and won the award Excellent main article award at the conference.
Building on Chinchilla and our NFNets and Perceiver multimodal models, we also present Flamingo, a family of visual language models for multi-shot learning. By handling images, videos and text data, Flamingo bridges the gap between vision-based and language-only models. Flamingo’s single model breaks the state of the art in multi-step learning across a wide range of open-ended multimodal tasks.
However, scale and architecture are not the only factors essential to the power of transformer-based models. Data properties also play an essential role, as we write about in the presentation on: data properties that promote contextual learning in transformer models.
Optimizing reinforcement learning
Reinforcement learning (RL) has shown great promise as an approach to creating generalized artificial intelligence systems that can tackle a wide range of intricate tasks. This has led to breakthroughs in everything from Go to math, and we’re always looking for ways to make RL agents smarter and more agile.
We introduce a fresh approach that enhances the decision-making capabilities of RL agents in a computationally capable manner drastically increasing the scale of information available to search for them.
We will also present a conceptually basic but general approach to curiosity-driven exploration in visually intricate environments – an RL agent called BYOL-Discover. It achieves superhuman performance while being immune to noise and is much simpler than previous works.
Algorithmic progress
From compressing data to performing weather prediction simulations, algorithms are a fundamental part of newfangled computing. That’s why incremental improvements can have a huge impact when working on a gigantic scale, helping you save energy, time and money.
We provide a radically fresh and highly scalable method automatic configuration of computer networksbased on neural algorithmic reasoning, showing that our highly versatile approach is up to 490 times faster than the current state of the art while meeting most input constraints.
In the same session, we also present a demanding exploration of the previously theoretical concept of “algorithmic matching”, highlighting the complex relationship between graph neural networks and dynamic programmingand how best to combine them to optimize performance outside of distribution.
Being a pioneer responsibly
At the heart of DeepMind’s mission is our commitment to act as responsible pioneers in artificial intelligence. We are committed to developing artificial intelligence systems that are lucid, ethical and fair.
Explaining and understanding the behavior of intricate AI systems is an necessary part of creating fair, lucid and precise systems. we offer a set of desiderata that reflect these ambitions and describe the practical way to achieve themwhich involves training an artificial intelligence system to build a causal model of itself, enabling it to explain its own behavior in a meaningful way.
To operate safely and ethically in the world, AI agents must be able to justify harm and avoid harmful actions. We will present collaborative work on a fresh statistical measure called counterfactual harmand demonstrate how it solves problems using a standard approach to avoid harmful policies.
Finally, we present our fresh article which proposes ways to diagnose and mitigate model reliability errors caused by changes in distributionshowing how essential these issues are to implementing unthreatening learning technologies in healthcare settings.
See the full scope of our work at NeurIPS 2022 Here.