Google DeepMind at NeurIPS 2023

Towards more multimodal, powerful and public artificial intelligence systems

Next, the beginning of the thirty -seventh annual conference of neuphs, the largest artificial intelligence conference (AI) in the world. NeuPIPs 2023 will be held from 10 to 16 December in New Orleans, USA.

Google DeepMind teams offer more than 180 sheets at the main conference and workshops.

We will display the illustrations of our advanced models of artificial intelligence to predict global weather and discover the materials and content created by AI. There will also be an opportunity to listen to the team behind the Gueini, which is the largest and most AI.

Here is a look at some of our most prominent research:

Artificial intelligence models can create paintings, compose music, and write stories. But although these models may be in one way, most of them struggle to transfer these skills to another. We delve into how to help obstetric capacity in the methods. In a presentation of the lights, we appear that spread models can be used to classify images without any additional required training. Classification of prevalence models such as IMAGEN images in a more like a person similar to other models, and relying on shapes instead of texture. What’s more, we explain how only a prediction can be made with illustrations of images improve vision learning. Our approach has exceeded the current methods of vision and tasks, and showed more capabilities to expand.

More multimedia models can give the most useful digital aids and robotin to help people in their daily lives. In Spotlight, the Wecreate factors that can interact with the digital world as humans do – through screen shots, keyboard and mouse verbs. Separately, we explain that by taking advantage of video generation, including translation and closed designation, models can transfer knowledge by predicting video plans for real robot procedures.

One of the following attractions can be to generate realistic experience in response to the procedures carried out by humans, robots and other types of interactive factors. We will show an experimental offer for UNISIM, our global simulation of interactions in the real world. This type of technology can have applications across industries from video and movies, to the real world’s training agents.

When developing and publishing large models, the privacy should be included in each step on the road.

In a recognized paper with Neups Best Paper, researchers explain how to evaluate training that preserves privacy with effective technology enough to use in the real world. For training, our teams study how to measure whether language models keep data – in order to protect special and sensitive materials. In another oral presentation, our scientists are looking for training restrictions through “student” and “teachers” models that have different levels of access and weakness if the attack is done.

Big language models can generate impressive answers, but they are vulnerable to “hallucinations”, the text that looks correct but made. Our researchers arouse our question whether the method of finding a stored (resettlement) site can enable the release of the truth. Surprisingly, they found that modifying the reality and editing the site does not liberate the truth, which hints to the complexity of understanding and controlling the information stored in LLMS. With TracR, we suggest a new way to evaluate the methods of interpretation by translating human readable programs into transformer models. We have opened a TRARR version to help work as a ground governorate to assess methods of interpretation.

Since large models become more capable, our research pushes the boundaries of the new capabilities to develop the most general artificial intelligence systems.

While language models are used for general tasks, they lack an exploratory understanding and the context necessary to solve the most complex problems. We offer the tree of ideas, a new framework to infer the language model to help models explore a wide range of possible solutions. By organizing logic and planning as a tree instead of the commonly used flat chain, we explain that the language model is able to solve complex tasks such as “Game 24” more accurately.

To help people solve problems and find what they are looking for, artificial intelligence models need to treat billions of unique values ​​efficiently. With the multiplicity of features, one representation space is used for many different features, allowing the large inclusion moderation (LeMS) to expand the range of products for billions of users.

Finally, with Doremi, we explain how artificial intelligence to automate the mixture of training types can greatly accelerate the language model training and improve performance on new and invisible tasks.

Enhancing the global artificial intelligence community

We are proud to sponsor NeuPIPs, support Latinx workshops in AI and Queerinai and women in ML, which helps to enhance research cooperation and develop a society of artificial intelligence and machine learning diverse. This year, NeuPs will have a creative path characterized by our visualization project, which creates committees to create more varied and accessible representations of artificial intelligence.

If you are preparing Neurips, come before our knees to learn more about our advanced research and meet workshops to host our teams and present them all over the conference.