Developing the best large models in the class, RL computer agents and more transparent, ethical and honest AI systems
Thirty -sixth international conference on neural information processing systems (Neuroips 2022) takes place from November 28 – December 9, 2022 as a hybrid event based in Nowy Orleana, the USA.
Neurips is the world's largest conference in the field of artificial intelligence (AI) and machine learning (ML) and we are proud that we can support this event as diamond sponsors, helping to support the exchange of research progress in the AI and ML community.
Teams from all over Deepmind present 47 articles, including 35 external cooperation in virtual panels and poster sessions. Here is a short introduction to some studies, which we present:
The best large models in the classroom
Large models (LMS) – generative AI systems trained on huge amounts of data – resulted in amazing performances in areas, including language, text, sound and image generation. Part of their success depends on their scale.
However, we created in Chinchilla Parameter language model of 70 billion, which exceeds many larger modelsincluding Gopher. We updated the rules on the scaling of large models, showing how previously trained models were too large to get the amount of training performed. This work has already shaped other models that follow these updated rules, creating slimmer, better models and won Excellent main paper Award at the conference.
We also present that building on chinchillas and our multimodal models Flamingo, a family with slight learning of visual language models. Operation of images, films and text data, Flamingo represents a bridge between vision models and only for language models. A single Flamingo model sets a new most modern learning in a small number of shots in a wide range of open multimodal tasks.
And yet the scale and architecture are not the only factors that are important for the power of models based on the transformer. Data properties also play an important role, which we discuss in the presentation Data properties that promote learning in context in transformer models.
Optimization of reinforcement learning
Learning (RL) turned out to be a great promise as an approach to creating generalized AI systems that can solve a wide range of complex tasks. This led to breakthroughs in many domains from mathematics and we are always looking for ways to make RL agents smarter and slimmer.
We are introducing a new approach that increases the decision -making skills of RL agents in a calculation way drastically expanding the scale of information available for recovery.
We will also present a simple conceptual, but a general approach to an exploration based on curiosity in visually complex RL agent-called environments called Byol-Explore. It achieves superhuman performance, at the same time solid for noise and is much simpler than previous work.
Algorithmic progress
From data compression to starting the simulation to predict weather, algorithms are the basic part of modern calculations. And so incremental improvements can have a huge impact when working on a scale, helping to save energy, time and money.
We radically divide a 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 art, while meeting most of the input restrictions.
During the same session, we also present a rigorous study of the previously theoretical concept of “algorithmic alignment”, Emphasizing the nuance relationship between graph neural networks and dynamic programmingAnd how to best combine them to optimize performance outside the distribution.
Pioneering responsibly
The heart of the Deepmind mission is our commitment to act as responsible pioneers in the field of AI. We are involved in developing AI systems that are transparent, ethical and honest.
Explanation and understanding of the behavior of complex AI systems is an important part of the creation of honest, transparent and accurate systems. We offer The designer set that captured these ambitions and describe a practical way to meet themwhich includes the training of the AI system to build a causal model, enabling it to explain his own behavior.
To act safely and ethically in the world, AI agents must be able to reason harm and avoid harmful actions. We will introduce cooperation on an innovative statistical measure called alternative harmAnd show how he overcomes problems with standard approaches to avoid harmful politicians.
Finally, we present our new article that I propose Ways of diagnosing and alleviating failures in the justice of the model caused by distribution changesshowing how important these problems are for implementing safe ML technologies in healthcare conditions.
See the full range of our work in Neuips 2022 Here.
