Google DeepMind Unveils Cutting-Edge AI Research at NeurIPS 2023



Google DeepMind Unveils Cutting-Edge AI Research at NeurIPS 2023

New Orleans, LA (December 8, 2023) - Google DeepMind, a leader in artificial intelligence research, showcased its latest advancements at the Neural Information Processing Systems (NeurIPS) conference, the largest AI conference globally. With over 180 research papers presented across the main conference and workshops, DeepMind demonstrated its commitment to pushing the boundaries of AI research and development.

Multimodality: Bridging the Gap Between Language, Video, and Action

DeepMind's research delves into the potential of multimodality, exploring how models can learn across different modalities like language, video, and action. This includes:

UniSim: A universal simulator of real-world interactions, paving the way for realistic experiences in response to human and robot actions.

Diffusion models: Classifying images in a more human-like way by relying on shapes rather than textures. Predicting captions from images can also improve computer vision learning.

Multimodal agents: Interacting with the digital world through screenshots, keyboard, and mouse actions, and transferring knowledge by predicting video plans for real robot actions.

These advancements have the potential to create more intuitive and helpful digital and robot assistants for everyday tasks.

Building Safe and Understandable AI

DeepMind also focuses on building safe and understandable AI. Key highlights include:

Large Language Models (LLMs): Investigating whether fact localization enables editing, revealing the complexity of understanding and controlling stored information.

Tracr: A novel interpretability evaluation method that translates human-readable programs into transformer models.

Privacy-preserving training: Techniques for measuring data memorization in LLMs and evaluating privacy-preserving training for real-world applications.

Student-teacher models: Exploring the limitations of training through models with different access levels and vulnerability.

These advancements aim to ensure that AI is developed and deployed responsibly, addressing concerns about safety, bias, and privacy.

Emergent Abilities: Pushing the Limits of AI Capability

DeepMind is also pushing the limits of AI capabilities:

Tree of Thoughts: A new framework for language model inference, improving exploration and reasoning for complex problem-solving.

Feature Multiplexing: Enabling large embedding models (LEMs) to scale to billions of users by using one representation space for many features.

DoReMi: Automating the mixture of training data types to significantly speed up language model training and improve performance on new tasks.

These advancements lay the groundwork for more general and capable AI systems that can tackle increasingly complex challenges.

Fostering a Global AI Community

DeepMind believes in fostering a diverse and inclusive AI community. The company sponsors workshops led by LatinX in AI, QueerInAI, and Women In ML, and participates in the NeurIPS Creative Track through its Visualising AI project. This commitment to diversity and inclusion is crucial for ensuring that AI benefits everyone.

Join Google DeepMind at NeurIPS 2023

To learn more about DeepMind's cutting-edge research and meet its team, visit the DeepMind booth at NeurIPS 2023. You can also explore the research papers and schedule on the DeepMind website.

Google DeepMind at NeurIPS 2023 schedule

Follow DeepMind for updates on the latest innovations

Sign up for the DeepMind newsletter


What is NeurIPS? The Neural Information Processing Systems (NeurIPS) is the largest conference on artificial intelligence and machine learning globally.

What is DeepMind? Google DeepMind is a research laboratory focused on artificial intelligence, with a mission to solve intelligence and use it to benefit humanity.

What are some of the key takeaways from DeepMind's research at NeurIPS 2023? DeepMind showcased advancements in multimodality, safe and understandable AI, and emergent abilities. The company is also committed to fostering a diverse and inclusive AI community.