engblogs

Perceiver AR is a general-purpose autoregressive generation model that can handle long-context inputs, outperforming standard Transformers and Transformer-XL models.

An autoregressive architecture - Perceiver AR - is developed for long-context density estimation via cross-attention and end-to-end causal masking without the use of hand-crafted sparsity patterns or memory mechanisms.

Paving the way for generalised systems with more effective and efficient AI

DeepMind's latest research presented at ICML 2022 covers advancements in machine learning across various fields.

Exploring the innovative capabilities of DALL·E 2 in expanding artistic horizons

Applying AI research to enhance YouTube experience and improve products and services used by billions of people every day.

An online math tutoring service leverages AI to enhance students' skills and confidence.

Providing PR templates as a guide to enhance the effectiveness of pull requests in code review process.

Instagram Engineering content will now be posted on the Engineering at Meta blog, allowing for more regular updates on the novel engineering work being done at Instagram.

Intuitive physics learning in a deep-learning model inspired by developmental psychology

Exploring the use of a deep-learning model inspired by developmental psychology to improve AI systems' understanding of intuitive physics

Analyzing the effectiveness of policy gradient methods in cooperative multi-agent reinforcement learning (MARL) and demonstrating how they can outperform value decomposition methods in certain scenarios, such as environments with highly multi-modal reward landscapes, with proper input representation and hyper-parameter tuning.