Research for AI models that can generalize, scale and accelerate learning
Next week starts 11. International conference on the representation of learning (ICLR), taking place on May 1-5 in Kigali, Rwanda. It will be the first main conference of artificial intelligence (AI) organized in Africa and the first personal event from the beginning of the pandemic.
Scientists from around the world will gather to share their most modern work in deep learning covering the fields of artificial intelligence, statistics and data as well as applications, including Machine Vision, Gaming and Robotics. We proudly support the conference as a sponsor of diamonds and master Dei.
Teams from all over Deepmind present 23 articles this year. Here are some attractions:
Open questions on the road to Aga
Recent progress has shown incredible AI performance in the text and image, but further research is needed to generalize systems between domains and scales. It will be a key step on the way to the development of artificial general intelligence (Aga) as a transformation tool in our daily lives.
We present a new approach in which models Learn by solving two problems in one. By training models in which you can look at the problem of two perspectives at the same time, learn to reason tasks that require solving similar problems, which is beneficial to generalize. We also examined The ability of neuronal networks to generalize Comparing them with Chomsky Language Hierarchy. Stranged testing of 2200 in 16 different tasks, we discovered that some models are fighting for generalization and we found that increasing them with external memory is crucial for improving performance.
Another challenge we are dealing with is how make progress in long -term tasks at the level of expertsWhere the prizes are not much and far away. We have developed a new set of data on Open Source approach and training to help models learn to discover in human ways for a long time.
Innovative approaches
When we develop more advanced possibilities of artificial intelligence, we must ensure that the current methods worked as intended and efficient for the real world. For example, although language models can bring impressive answers, many cannot explain their answers. We present Method of using language models to solve problems with multi -stage reasoning Using their basic logical structure, giving explanations that can be understood and checked by people. On the other hand, the opposite attacks are a way to study the boundaries of AI models by pushing them to create improper or harmful results. Training on the opposite examples makes the models more resistant to attacks, but they can cost performance on “regular” input data. We show this by adding adapters, we can create models that allow us to control this compromise in flight.
Learning (RL) turned out to be effective for the series Real challengesBut RL algorithms are usually designed to do one task well and try to generalize to new ones. We propose algorithm distillationA method that allows you to effectively generalize a single model for new tasks by training the transformer imitating the history of learning RL algorithms in various tasks. RL models also learn according to trials and errors that can be very demanding data and time consuming. Our model agent 57 took almost 80 billion data frame to achieve performance at a human level in 57 Atari games. We share a new way Train to this level using 200 times less experiencesignificantly reducing calculation costs and energy.
Ai for science
AI is a powerful tool for scientists to analyze huge amounts of complex data and understanding the world around us. Several articles show how AI accelerates scientific progress – and how science develops artificial intelligence.
Anticipating the properties of a 3D molecule is crucial for discovering drugs. We present Denoising method This achieves the newly most modern molecular properties in the forecast, enables a large -scale initial and generalizing in various biological data sets. We also present a new one Transformer that can make more accurate calculations of quantum chemistry Using data on atomic positions themselves.
Finally with FignetWe attract inspiration from physics to modeling collisions between complex shapes, such as a teapot or donut. This simulator can apply to robotics, graphics and mechanical design.
View Full List Deep -sea documents and schedule of events at ICLR 2023.
