DeepMind by Google at ICML 2024

Sure! Here’s the revised content with any promotional calls-to-action rephrased and external website references replaced:

Research

Published
19 July 2024

Exploring AGI, the challenges of scaling and the future of multimodal generative AI

Next week, the artificial intelligence (AI) community will convene for the 2024 International Conference on Machine Learning (ICML). Running from July 21-27 in Vienna, Austria, this conference serves as an international platform for showcasing the latest advancements, exchanging ideas, and shaping the future of AI research. This year, teams from Google DeepMind will present over 80 research papers. We will also present our multimodal on-device model, Gemini Nano, our new family of AI models for education called LearnLM, and demonstrate TacticAI, an AI assistant designed to assist with football tactics. Here, we introduce some of our oral, spotlight, and poster presentations:

Defining the path to AGI

What is artificial general intelligence (AGI)? This term refers to an AI system that matches human capabilities in a broad range of tasks. As AI models progress, outlining what AGI could entail in practical terms becomes increasingly crucial. We will offer a framework for classifying the capabilities and behaviors of AGI models, categorizing systems from non-AI calculators to emerging AI technologies based on performance, generality, and autonomy. Furthermore, we will illustrate the significance of open-endedness in establishing generalized AI that surpasses human capabilities. While recent AI advancements have largely leveraged existing Internet-scale data, open-ended systems can foster new insights that enhance human knowledge.

At ICML, we will be demoing Genie, a model capable of generating a variety of playable environments based on text prompts, images, photos, or sketches.

Scaling AI systems efficiently and responsibly

The development of larger, more proficient AI models necessitates more efficient training methods, alignment with human preferences, and improved privacy safeguards. We will demonstrate that employing classification instead of regression techniques facilitates the scaling of deep reinforcement learning systems, achieving state-of-the-art performance across various domains. Additionally, we propose a novel method for predicting the distribution of potential consequences from a reinforcement learning agent's actions, enabling rapid assessment of new scenarios. Our research team presents an alignment-maintaining strategy that reduces the need for human oversight, along with a new game theory-based approach to fine-tuning large language models (LLMs) that better aligns outputs with human preferences. We critique training models exclusively on public data and fine-tuning them with "differentially private" techniques, arguing that this method may not deliver the claimed levels of privacy or utility.

VideoPoet is a large language model for zero-shot video generation.

New approaches in generative AI and multimodality

Generative AI technologies and multimodal capabilities are broadening the creative horizons of digital media. We will showcase VideoPoet, which employs an LLM to create state-of-the-art video and audio from multimodal inputs such as images, text, audio, and other videos. We will also share Genie (generative interactive environments), which can produce a variety of playable environments to train AI agents based on text prompts, images, photos, or sketches. Lastly, we introduce MagicLens, an innovative image retrieval system that utilizes text instructions to fetch images that exhibit richer relationships beyond mere visual similarity.

Supporting the AI community

We are proud to sponsor ICML and promote a diverse community in AI and machine learning by backing initiatives led by Disability in AI, Queer in AI, LatinX in AI, and Women in Machine Learning. If you’re attending the conference, feel free to visit the Google DeepMind and Google Research booths to meet our teams, witness live demonstrations, and discover more about our research.

This version maintains the core information while ensuring a neutral tone and branding aligned with Hotnchill.