Fellowship: Genie, A Model for Generative Interactive Environments скачать в хорошем качестве

Fellowship: Genie, A Model for Generative Interactive Environments

#AI #GenerativeModels #UnsupervisedLearning #InteractiveEnvironments #OpenSourceAI #GitHub #arXiv Link to paper: https://arxiv.org/abs/2311.03054 Paper by: Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Bechtle, Feryal Behbahani, Stephanie Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando de Freitas, Satinder Singh, Tim Rocktäschel Presentation by the team at: https://www.fellowship.ai/ Fellowship.ai, part of Launchpad.ai, propels the forefront of AI research to power its programs. For a deep dive into our AI innovations and solutions designed for diverse business challenges, visit: https://www.launchpad.ai/ai-developme... The paper introduces "Genie", the first generative interactive environment trained unsupervised from unlabelled Internet videos. Comprising an 11B parameter model, Genie includes a spatiotemporal video tokenizer, an autoregressive dynamics model, and a scalable latent action model. This foundation world model enables interaction with generated environments on a frame-by-frame basis, without the necessity for ground-truth action labels or other domain-specific prerequisites commonly found in world model literature. Moreover, the learned latent action space of Genie facilitates the training of agents to mimic behaviors from unseen videos, paving the way for the development of versatile future agents. Explore more about this groundbreaking project here: https://github.com/tyxsspa/genie