Spotlight on Paris, London, Tel Aviv and Zurich
Eight years after establishing its FAIR hub in Paris, Meta has become one of the world’s leading research organizations, with pioneering research emerging from technology hubs in Paris, London, Tel Aviv and Zurich.
One of the most important decisions we made in founding FAIR was to center exploratory research and open science. We regularly collaborate with external researchers. This is because we have a strong hypothesis that this is the fastest and most responsible way to make progress.
“We have worked with various institutions to train generations of AI researchers, especially through our PhD programs,” said Naira Murray, Director of FAIR EMEA. “Many of our PhD students have made important contributions to this field.”
Our teams, currently located in Paris, London, Tel Aviv and Zurich, focus on Self-Supervised Learning, Reinforcement Learning, Speech and Speech, Computer Vision, Natural Language Modeling, Responsible AI, Machine Learning Theory, Model Efficiency, AR/VR Such.
“Our research is driven by a unique combination of ambition and camaraderie, with our teams working closely across boundaries of expertise, seniority, location and function to rapidly advance research. We do,” Murray said. “In this era of AI research, every day brings the possibility of new breakthroughs in research, including for his EMEA team at our company.”
Research on groundbreaking large-scale language models
Earlier this year, our researchers in Paris formed a team to build and deploy. llama (Large Language Model Meta AI) – State-of-the-art foundation large scale language model It is designed to help researchers advance their research in this sub-field of AI.
LLaMA works by taking a sequence of words as input and recursively generating text by predicting the next word. To train the model, we selected text from the 20 most spoken languages, with an emphasis on languages with Latin and Cyrillic scripts. With the ability to generate creative text, solve math theorems, Predict protein structureLarge-scale language models, such as answering reading comprehension questions, are one of the most obvious cases of the potential huge benefits that AI can offer at scale to billions of people.
Self-monitoring computer vision research
Our team, also based in Paris, has made two breakthroughs in computer vision research. Announced in April DINOv2 – A first method for training computer vision models that uses self-supervised learning to achieve results that meet or exceed standard approaches used in the field.
DINOv2 can detect and segment objects in images or videos without supervision or giving targets. For example, DINO can understand that an image contains a representation of a dog without being told what a dog is in the first place. As part of this announcement, we shared: public demo Anyone can use it to explore some of DINOv2’s features.
We are already using DINOv2 to learn more about the physical world. Meta recently World Resources Institute To Map forests with AI – tree by tree – over an area the size of a continent. Our self-supervised model was trained on data from forests, In North America, evaluations confirm that this generalizes well and provides accurate maps elsewhere in the world.
Our Paris team, in collaboration with our North American colleagues, has also pioneered new research using: SEER (SElf-SupERvised), a breakthrough self-supervised computer vision model from Meta AI Research. SEER directly learns from arbitrary random image collections and outputs image embeddings without the need for careful data curation and labeling required in traditional computer vision training.
Our latest breakthrough, SEER10B, uses diverse datasets for better and unbiased computer vision. Traditional computer vision systems are trained primarily on examples of wealthy countries in the United States and Europe, so they often do poorly on images from other locations with different socioeconomic characteristics. SEER provides excellent results for images from all over the world, including regions of various income levels outside the United States and Europe. SEER10B significantly improved fairness benchmark performance across gender, apparent skin color and age groups. Apart from improving performance on fairness benchmarks, the model has a good understanding of images around the world and locates them with unprecedented accuracy. We hope SEER will be a key building block as the AI community strives to build systems that work well for everyone.
Advances in 3D modeling
In August 2022, researchers from London and Paris open sourced the code. implicitron, A modular framework within the open source PyTorch3D library. Implictron uses neural implicit representations. This is a computer vision technology that can seamlessly combine real and virtual objects in augmented reality. It does not require a large amount of data for training, nor is it limited to a small number of viewpoints.
Implicitron learns a representation of a 3D object or scene using a sparse set of combined images of that object or scene from arbitrary viewpoints. Unlike his traditional 3D representations such as meshes and point clouds, this new approach represents objects as continuous functions, resulting in more accurate reconstruction of shapes, including complex geometries, and higher color reconstruction accuracy. becomes possible.
Image and video generation AI
Our team in Tel Aviv works closely on generative AI and has been at the forefront of Meta’s latest advances. In July 2022, Tel Aviv researchers and collaborators from around the world will is created A generative AI research model called Make-A-Scene. This multimodal generative AI technique puts creative control in the hands of those who use it by allowing them to describe and illustrate visions through both textual descriptions and free-form sketches, resulting in Surreal art such as flying hot dogs and city skyscrapers are born. desert.
We followed up on this effort with make a videoan AI system that can turn text prompts into concise, high-quality, unique video clips. The system can also create videos from images or take existing videos and create new similar videos.
Metaverse and Beyond
W.We believe that augmented and virtual reality combined with AI-powered interfaces constitute the next paradigm shift in human-oriented computing. While his other EMEA hubs are primarily focused on AI research to help him reach his goals, his team in Zurich is working closely on his AR and VR advancements.
Together, we are working to develop contextualized AI interfaces that allow devices to understand our context, preferences, history, and goals. This supports our vision of the future where devices act as partners rather than tools, surrounding us with technology that adapts to us and helps us work the way we want.
Historically, various areas of AI research have been relatively isolated from each other, Murray says. However, the collaborative infrastructure that FAIR has built has become an important catalyst for bringing various teams together and advancing research.
Murray, head of the FAIR EMEA team, says one of the best parts of his job is “facilitating collaboration among researchers by pointing out connections between related research interests.” I said.
“Over the past few months, we’ve seen an exciting blend of multimodal perception, language understanding and generation, reinforcement learning, and human-machine interaction,” Murray said. “We are very excited that this merger brings us closer to the field’s long-held dream of building truly advanced intelligent systems.”
