DeepMind’s latest research at ICLR 2023

Research in artificial intelligence models that can generalize, expand and accelerate science

Next week, the beginning of the eleventh international conference to represent the ICLR, which will be held 1-5 in Kigali, Rwanda. This will be the first major conference of artificial intelligence (AI) to be hosted in Africa and the first personal event since the beginning of the epidemic.

Researchers from all over the world will meet to exchange their advanced work in deep learning that extends to the areas of artificial intelligence, statistics, data science and applications, including the vision of machines, games and robots. We are proud to support the conference as a diamond sponsor and a Dei hero.

Details from all over DeepMind offer 23 sheets this year. Here are some prominent points:

Open the road questions to AGI

The latest progress showed the amazing artificial intelligence performance in the text and image, but more research for systems is needed for generalization through fields and standards. This will be a decisive step on the road to developing artificial intelligence (AGI) as a transformational tool in our daily life.

We present a new approach where you learn models by solving two problems in one. By training models to consider a problem from two perspectives at the same time, they learn how to think about tasks that require solving similar problems, which is useful for generalization. We also discovered the ability of nerve networks to generalize by comparing them to the hierarchy of Chomsky languages. By testing 2200 models in 16 different tasks, we discovered that some models are fighting for generalization, and we found that enhancing them with external memory is very important to improve performance.

The other challenge we face is how to make progress in long -term tasks at the level of experts, where the rewards are few and far apart. We have developed a new approach and an open source training collection to help models learn exploration in human -like ways.

Innovative approaches

While developing more advanced capabilities of artificial intelligence, we must ensure that current methods work intended and efficiently for the real world. For example, although language models can produce impressive answers, many cannot explain their responses. We offer a way to use language models to solve multiple -step thinking problems by exploiting its basic logical structure, and providing interpretations that can be understood and verified by humans. On the other hand, aggressive attacks are a way to explore the limits of artificial intelligence models by pushing them to create wrong or harmful outputs. Training examples of litigation makes models more powerful for attacks, but can come at the cost of performance on “regular” inputs. We explain that by adding adapters, we can create models that allow us to control this barter while flying.

Reinforcement learning (RL) has proven successful to a set of challenges in the real world, but RL algorithms are usually designed to do one task well and struggle to generalize new transfers. We suggest the distillation of algorithm, a method that enables a single model of generalizing efficiently to new tasks by training an adapter to imitate the history of RL algorithms through various tasks. RL models also learn by experiment and error, which can be very dense of data and take a long time. It took approximately 80 billion frameworks of data for our 57 model to reach human level performance across 57 Atari games. We share a new method of training at this level using a 200 -time experience, which greatly reduces computing and energy costs.

Artificial Intelligence of Science

AI is a powerful tool for researchers to analyze huge amounts of complex data and understand the world around us. Many papers show how artificial intelligence accelerates scientific progress – and how science progresses from artificial intelligence.

The prediction of the characteristics of the molecule of its 3D structure is crucial to the discovery of drugs. We offer a way to reduce it to achieve the latest latest prediction of molecular property, allow pre -training on a large scale, and is circulated through different biological data groups. We also offer a new adapter that can make quantum chemistry accounts more accurate using data about atomic places alone.

Finally, with Fignet, we are inspired by physics to a collision design between complex shapes, such as tea jug or Donat. This simulation can have applications via robots, graphics and mechanical design.

See the full menu of DeepMind and the events schedule at ICLR 2023.