Discovering aga, scaling challenges and the future of multimodal generative AI
Next week, the artificial intelligence community (AI) will meet at 2024 International conference on machine learning (ICML). The conference on July 21-27 in Vienna in Austria is an international platform for presenting the latest achievements, exchange of ideas and shaping the future of AI research.
This year, teams from all over Google Deepmind will present over 80 research articles. At our stand we will also present our multimodal model on the device, Gemini Nano, our new family of AI models for education called Learnlm And the demo of Tacticai, an AI assistant who can help in football tactics.
Here we present some of our oral presentations, headlights and posters:
Defining the path to Aga
What is artificial general intelligence (agi)? The expression describes the AI system, which is at least as talented as a man in most tasks. Because AI models are still developing, determining how Agi can look in practice will become more and more important.
We will present a framework for Classifying the possibilities and behavior of Agi models. Depending on their performance, generality and autonomy, our article categorizes systems, from calculators other than AI to appearing AI models and other new technologies.
We will also show Openness is crucial for building generalized artificial intelligence It goes beyond human possibilities. While many of the last AI achievements were based on existing data on the internet scale, open systems can generate new discoveries that expand human knowledge.
In ICML we will demonstrate the gene, a model that can generate a number of played environments based on text hints, images, photos or sketches.
Scaling AI systems efficiently and responsibly
The development of larger, more talented AI models requires more efficient training methods, closer adaptation to human preferences and better privacy security.
We will show how to use Classification instead of regression techniques It facilitates the scaling of deep reinforcement learning systems and achieving the latest performance in various domains. In addition, we offer a new approach provides for the distribution of the consequences of the reinforcement agent agent activitieshelping to quickly evaluate new scenarios.
Our researchers present approach to maintenance This reduces the need for human supervision and New approach to tuning large language models (LLM)Based on the theory of games, it better adapts the LLM result to human preferences.
We Critate the approach of training models on public data and only tuning the “Private” trainingAnd they argue that this approach may not offer privacy or usability, which he often claims.
Videopoet is a large language model to generate zero video.
New approaches in generative artificial intelligence and multimodality
AI generative technologies and multimodal possibilities expand the creative capabilities of digital media.
We will present Videoetwhich uses LLM to generate the most modern videos and sound from multimodal inputs, including images, text, sound and other videos.
And divide Gin (Generative Interactive Environment), which can generate a number of playable environments for training AI agents, based on text hints, images, photos or sketches.
Finally we present MagiclensAn innovative image download system that uses text instructions to download images with richer relationships outside the visual similarity.
Supporting the AI community
We are proud of sponsoring ICML and supporting various communities in artificial intelligence and machine learning by supporting initiatives conducted by Disability in artificial intelligenceINQueer in AIINLatinx in AI ANDMachine learning women.
If you are at a conference, visit Google Deepmind and Google Research stands to get to know our teams, see the demo live and learn more about our research.
