engblogs

BYOL-Explore is a curiosity-driven AI agent that learns a world representation, world dynamics, and exploration policy together by optimizing prediction error at the representation level.

A blogpost on BYOL-Explore, a method for curiosity-driven exploration in visually-complex environments by optimizing a single prediction loss in the latent space without auxiliary objectives.

Exploring advanced machine learning techniques with large models

Unlocking high-accuracy differentially private image classification through scale.

Exploring the impact of large models on the evolution of technology

Investigating and improving high-accuracy differentially private image classification through scale adjustments and modifications to training procedure and model architecture.

Bridging DeepMind research with Alphabet products, exploring the impact of vision language models and the potential uses in various fields.

Connecting DeepMind research with Alphabet products through vision language models at the AI Summit

Leveraging AI technology to enhance human error detection through critiques.

Training AI models to critique flaws in summaries helps humans notice them more often, highlighting the potential for AI systems to assist with supervising difficult tasks.

Optimizing training efficiency for large neural networks

Training large neural networks requires orchestrating a cluster of GPUs to perform a single synchronized calculation.