Towards more multimodal, solid and general AI systems
Next week is the beginning of the 37th annual conference on neural information processing systems (Neuips), the largest artificial intelligence conference (AI) in the world. Neuroips 2023 It will take place on December 10-16 in Nowy Orleana, the USA.
Teams from all over Google Deepmind present over 180 articles at the main conference and workshops.
We will present versions of our most modern AI models for global weather forecasting, discovering materials and a water marking generated by AI. It will also be possible to hear from the team for Gemini, our largest and most talented AI model.
Here are some of our most important research points:
Multimodality: language, video, action
Unisim is a universal simulator of interaction in the real world.
AI generative models can create images, compose music and write stories. But although talented these models can be in one medium, most fight to transfer these skills to others. We delve into how generative skills can help in learning between methods. In the presentation in the spotlight we show this Diffusion models can be used to classify images No additional training. Dyfusion models, such as imagen, classify images in a more human way than other models, based on shapes rather than textures. What's more, we show how fair Predicting signatures from images can improve computers learning. Our approach exceeded the current methods of view and language tasks and showed greater potential for scaling.
More multimodal models can give way to more useful digital assistants and robots to help people in their daily lives. In the reflector's poster, weCreate agents who could interact with the digital world, just like people – through screenshots and keyboard and mouse actions. We show it separately by Using video generation, including inscriptions and closed signatures, models can transfer knowledge Anticipating video plans of real robot activities.
One of the next milestone may be generating realistic experience in response to activities carried out by people, robots and other types of interactive agents. We will present a demo UnisimOur universal interaction simulator in the real world. This type of technology can be used in various industries, from video games and film, to training agents for the real world.
Building safe and understandable artificial intelligence
Illustration of the artist's artificial intelligence (AI). This image presents AI safety research. It was created by the artist Khyati Trehan as part of the AI visualization project launched by Google Deepmind.
When developing and implementing large models, privacy must be embedded at every stage.
In an article recognized with Neurips Best Paper AwardOur researchers show how to assess privacy protection Training with effective technique Just for use in the real world. For the training, our teams study how to measure if Language models remember data – to protect private and sensitive materials. In another oral presentation, our scientists study Restrictions on training through models of “students” and “teachers” which have different levels of access and susceptibility if they are attacked.
Large language models can generate impressive answers, but are susceptible to “hallucinations”, a text that seems correct, but is invented. Our researchers ask the question whether the method of finding a location stored by facts (location) may allow you to edit the fact. They found this surprisingThe location of the fact and edition of the location does not edit the factPointing to the complexity of understanding and controlling stored information in LLM. WITH Tracr, we suggest a new way to assess the interpretation Methods by translating human programs into transformer models. We have Open the TRACR version obtained To help as difficult to assess interpretation methods.
Rising abilities
Illustration of the artist's artificial intelligence (AI). This image imagines artificial general intelligence (Agi). It was created by Novoto Studio as part of the AI visualization project launched by Google Deepmind.
Because large models become more capable, our research exceeds the boundaries of new development skills to develop more general AI systems.
While language models are used for general tasks, they lack the necessary exploration and contextual understanding to solve more complex problems. We present Tree of Thought, a new language model inference To help explore models and reason in a wide range of possible solutions. By organizing reasoning and planning as a tree instead of a commonly used flat chain, we show that the language model is able to solve complex tasks, such as “Game 24” much more accurately.
To help people solve problems and find what they are looking for, AI models must effectively process billions of unique values. WITH Multiplexing of the function, one representative space is used for many different functionsenabling large embedding models (LEM) for scaling products for billions of users.
Finally, with Doremi, we show how the use of artificial intelligence for automation A mixture of training types can significantly accelerate the training of the language modeland improve the performance of new and invisible tasks.
Supporting the global AI community
We are proud of sponsoring Neurips and supporting workshops conducted by Latinx in AIIN QueerinaiAND Women in MLHelping supporting research cooperation and the development of a variety of AI community and machine learning. This year, Neuryps will have a creative song with our visualizing AI project, which provides artists to create more diverse and available representations of artificial intelligence.
If you attend neuripa, come with our stand to learn more about our most modern research and meet our teams organizing workshops and presenting themselves at the conference.
