Facebook AI year in review: 2019

January 08, 2020

Written byJerome Pesenti

Written by

Jerome Pesenti

Share

Something Went Wrong
We're having trouble playing this video.

As we start the new year, we wanted to look back at some of our most notable work in 2019. On the Facebook AI blog, we highlighted groundbreaking work that advanced the state of the art in self-supervised learning; cross-lingual and conversational language understanding; hidden-information games like poker and Hanabi; 3D content understanding; medical imaging; and more.

Beyond our pure research efforts, Facebook AI also shared a wide range of new resources and tools to help the greater AI research community. And finally, we discussed how Facebook is using AI to create new experiences and better serve the people who use our products.

We still have much work ahead to achieve our ultimate goal of surpassing human intelligence. But I’m excited by the progress we’ve made and proud of our open and collaborative approach. Here are some highlights from the past year.

Research

Facebook AI pushed the field in fresh directions in 2019, and found new ways to use AI to make a positive impact on the world and ensure that the technology is used responsibly. Our researchers saw notable successes in self-supervised learning, which we believe will be important to the next generation of AI breakthroughs. They also opened new horizons for deep learning in areas such as medicine, biology, fashion, and symbolic math.

Our researchers were also honored with several notable awards, including the Turing Award, presented to our Chief AI Scientist, Yann LeCun, and the Steven A. Coons Award, presented to our Director of Computational Photography, Michael Cohen. In addition to the many papers we published, Facebook AI also organized and participated in open challenges, such as IntPhys and Zero Speech. Here are some of our blog posts about our noteworthy research projects:

Mapping the world to help aid workers, with weakly-, semi-supervised learning
Facebook, Carnegie Mellon build first AI that beats pros in 6-player poker
RoBERTa: An optimized method for pretraining self-supervised NLP systems
Facebook AI leads in 2019 WMT international machine translation competition
Advances in conversational AI
Building AI to inform people’s fashion choices
Pushing the state of the art in 3D content understanding
XLM-R: State-of-the-art cross-lingual understanding through self-supervision
Results of the first fastMRI image recognition challenge
Building AI that can master complex cooperative games with hidden information

Resources

We have a core belief in the ability of open science, shared resources, and reproducible research to accelerate the pace of AI research and engineering. Here are some of the tools and resources we shared last year:

PyTorch adds new dev tools as it hits production scale
Open-sourcing AI Habitat, a simulation platform for embodied AI research
PyTorch 1.2 release and Global Summer Hackathon
PyTorch 1.3 adds mobile, privacy, quantization, and named tensors
Detectron2: A PyTorch-based modular object detection library
Open-sourcing mvfst-rl, a research platform for managing network congestion with reinforcement learning

Responsible AI

In our research, in the technology utilized here at Facebook, and in the tools we share with the AI community, we are focusing on ensuring that AI is used responsibly. Here are some of the projects we discussed on the Facebook AI blog in 2019:

CrypTen: A new research tool for secure machine learning with PyTorch New progress in using AI to detect harmful content
How the AI community can get serious about reproducibility
Deepfake Detection Challenge launches with new data set and Kaggle site

AI-powered applications

In 2019, we shared details on how AI is powering a wide range of products and initiatives at Facebook. These include new AI-powered hardware, like Portal’s Smart Camera and the Oculus Insight system used in our latest VR devices. We also shared how we are using self-supervised learning, multimodal content understanding, and other techniques to keep people safe on our platforms.

Under the hood: Portal’s Smart Camera
Advancing self-supervision, CV, NLP to keep our platforms safe
Powered by AI: Oculus Insight
Using deep neural networks for accurate hand-tracking on Oculus Quest
Powered by AI: Instagram’s Explore recommender system

Written by

Jerome Pesenti

VP of Artificial Intelligence, Facebook