Advancing better large models, computationally optimal RL agents, and more transparent, ethical, and fair AI systems
The Thirty-sixth International Conference on Neural Information Processing Systems (NeuroIPS 2022) will take place from November 28 to December 9, 2022, as a hybrid event, based in New Orleans, United States.
NeurIPS is the world's largest conference on artificial intelligence (AI) and machine learning (ML), and we are proud to support the event as a Diamond Sponsor, helping to foster the exchange of advances in research in the AI and ML community.
DeepMind teams present 47 papers, including 35 external collaborations through virtual panels and poster sessions. Here is a brief introduction to some of the research we present:
The best large models in their category
Large models (LMs) – generative AI systems trained on enormous amounts of data – have resulted in incredible performance in areas such as language, text, audio and image generation. Part of their success is their scale.
However, at Chinchilla we have created a 70 billion parameter language model that outperforms many larger models, including Gopher. We updated the scaling laws for large models, showing how previously trained models were too large for the amount of training done. This work has already shaped other models that follow these updated rules, creating simpler and better models, and won a Outstanding main track paper prices at the conference.
Building on Chinchilla and our multimodal models NFNets and Perceiver, we also present Flamingo, a family of visual language models for learning in just a few steps. Handling images, videos, and text data, Flamingo represents a bridge between vision-only models and language-only models. A single Flamingo model establishes a new state of the art in few-step learning on a wide range of open-ended multimodal tasks.
And yet, scale and architecture are not the only important factors for the power of transformer-based models. Data properties also play an important role, which we discuss in a presentation on Data properties that promote in-context learning in transformer models.
Optimizing reinforcement learning
Reinforcement learning (RL) has shown great promise as an approach to creating generalized AI systems capable of handling a wide range of complex tasks. This has led to advances in many areas, from Go to mathematics, and we are always looking for ways to make RL agents smarter and simpler.
We introduce a novel approach that improves the decision-making capabilities of RL agents in a computationally efficient manner. greatly expanding the scale of information available for retrieval.
We will also present a conceptually simple but general approach for curiosity-driven exploration in visually complex environments: an RL agent called BYOL-Explorer. It achieves superhuman performance while being robust to noise and much simpler than previous work.
Algorithmic advances
From compressing data to running simulations to predict the weather, algorithms are a fundamental part of modern computing. So, incremental improvements can have a huge impact when working at scale, saving energy, time and money.
We share a radically new and highly scalable method for automatic configuration of computer networksbased on neural algorithmic reasoning, showing that our highly flexible approach is up to 490 times faster than the current state of the art, while satisfying the majority of input constraints.
In the same session, we also present a rigorous exploration of the previously theoretical notion of “algorithmic alignment”, highlighting the nuanced relationship between graph neural networks and dynamic programmingand how best to combine them to optimize non-distribution performance.
Pioneer responsibly
At the heart of DeepMind's mission is our commitment to acting as a responsible pioneer in the field of AI. We are committed to developing transparent, ethical and fair AI systems.
Explaining and understanding the behavior of complex AI systems is an essential part of creating fair, transparent and accurate systems. We offer a set of desiderata that capture these ambitions and describe a practical way to achieve themwhich involves training an AI system to build a causal model of itself, allowing it to explain its own behavior in a meaningful way.
To act safely and ethically in the world, AI agents must be able to reason about dangers and avoid harmful actions. We will present collaborative work on a new statistical measure called counterfactual harmand demonstrate how it overcomes problems with standard approaches to avoid pursuing harmful policies.
Finally, we present our new article which proposes ways to diagnose and mitigate model fairness failures caused by distribution changesshowing how important these questions are for the deployment of secure ML technologies in healthcare settings.
Discover all of our work at NeurIPS 2022 here.