Meta FAIR Shares New AI Research, Code, Models And Datasets

Over the past few months, Meta and its CEO Mark Zuckerberg have reiterated on many occasions that they believe an open-source environment is the best way forward to advance the artificial intelligence (AI) and generative artificial intelligence (GenAI) world.

The idea is to benefit all.

This comes despite some of its biggest competitors believing otherwise, especially looking at OpenAI, a company that promised to remain a non-profit, transforming itself into a for-profit venture.

As such, Meta has taken it to heart to share its research, products, code, and more in an open-source format, which has even led to a team dedicated to this cause.

Called The Fundamental AI Research or FAIR, this team works to further Meta’s “fundamental understanding in both new and existing domains, covering the full spectrum of topics related to AI, with the mission of advancing the state-of-the-art of AI through open research for the benefit of all.”

Over time, Meta has released a series of products, tools, and research materials that can be accessed by anyone to further their AI ambitions.

Building on that, the company revealed a new set of updates from FAIR in a blog post, which included its latest research, code, models, and datasets.

These updates are aimed at building more capable agents, boosting safety, and enabling architecture innovations that enhance how models learn new information.

Meta Motivo

Meta Motivo brings in a first-of-its-kind behavioral foundation model that controls virtual humanoid agents for complex tasks. It addresses commonly seen body control problems in digital avatars, allowing them to possess more human-like movements.

It uses an innovative algorithm to learn human-like behaviors from unlabeled motion data and enables the tool to handle tasks like motion tracking, pose reaching, and reward optimization without extra training. Its unique design embeds states, motions, and rewards into a shared latent space, ensuring efficient and adaptable performance.

Meta Motivo outperforms other methods in creating realistic behaviors and remains robust to environmental changes like gravity or wind.

Essentially this technology could revolutionize character animation and immersive experiences in the Metaverse, bringing lifelike virtual agents to life.

Meta Video Seal

Building on the boost for character animation and immersive experiences, Meta revealed Meta Video Seal, a state-of-the-art comprehensive framework for neural video watermarking.

This tool embeds a watermark into videos, one that can’t be seen by the naked eye. The watermark can also be accompanied by a hidden message. These watermarks can be later uncovered to determine the origins of a video.

It helps circumvent common workarounds such as blurring, cropping, compression, and other methods to enable watermark erasing.

To this effect, Meta is sharing the model under a permissive license, along with training and inference code, a research paper, and an interactive demo.

Meta also launched Omni Seal Bench, a leaderboard for neural watermarking research, and re-released its Watermark Anything model.

Source

Flow Matching

Flow Matching is a powerful GenAI framework designed to create diverse outputs like images, videos, audio, music, and 3D structures.

This method has replaced traditional diffusion methods in various Meta applications, including Meta Movie Gen, Meta Audiobox, and Meta Melody Flow, as well as industry projects like Stable-Diffusion-3 and more. Flow Matching offers efficiency and adaptability and simplifies generalizing complex data.

Meta has publicly released a paper, code, and training scripts for both continuous and discrete Flow Matching, enabling researchers to explore and build on this method. This initiative aims to encourage broader adoption of Flow Matching in generative AI projects worldwide.

Meta Explore Theory-of-Mind

Meta Explore Theory-of-Mind brings in a new approach to advancing social intelligence in AI, which is done by focusing on reasoning about others' thoughts and beliefs.

Unlike existing datasets that are limited to evaluation and narrow scenarios, this framework generates diverse, challenging, and scalable data for training and testing Theory-of-Mind (ToM) reasoning. By enabling richer interactions and more robust ToM datasets, Meta aims to accelerate progress in this crucial area, paving the way for more advanced machine intelligence.

“Explore Theory-of-Mind generates robust and reliable stories that push the limits of large language models (LLMs), making it ideal for evaluating frontier models or fine-tuning data.”

Large Concept Models

Meta’s Large Concept Model (LCM) introduces a new language modeling paradigm by separating reasoning from language representation.

Behind the idea for this technology lies how humans plan ideas before communicating. As such, LCM predicts high-level concepts, like sentences, in a multimodal and multilingual embedding space, rather than just predicting the next word.

By sharing this research, Meta aims to advance language models capable of operating across languages and modalities in a structured, hierarchical way.

Image Diversity Modeling

Working to advance the safe development of image generation models with new research and tools, Meta’s FAIR announced an updated model that builds on earlier work to prioritize realistic, high-quality images while ensuring responsibility and safety.

To further push progress in this area, the company is collaborating with external experts to research methods for responsible image diversity modeling, while also open-sourcing a comprehensive evaluation toolbox for text-to-image generative models.

Other Releases

Meta released its Dynamic Byte Latent Transformer, which is a tokenizer-free language model that processes text at the byte level, improving efficiency and robustness for long and rare sequences. It outperforms traditional models and advances capabilities in areas like low-resource languages, coding, and factual reasoning.

Meta’s Memory Layers introduces a scalable method to enhance factuality in LLMs by using trainable key-value lookups to store and retrieve information efficiently. This approach brings greater performance with lower computational demands while outperforming dense and mixture-of-expert models on downstream tasks.

Meta CLIP 1.2 offers a cutting-edge vision-language encoder designed to map image and text semantics with precision using curated, high-quality datasets, which aims to empower researchers and developers to advance vision-language understanding and build large-scale datasets for diverse applications.

“This work supports our long and proven track record of sharing open reproducible science with the community,” said Meta in the blog post announcing the release of the new artifacts by FAIR. “By publicly sharing our early research work, we hope to inspire iterations and ultimately help advance AI in a responsible way.”

What do you think of Meta’s new releases? Do you think Meta’s open-source environment is the way forward for AI and generative AI technologies?

Let us know in the comments below!

First published on Fri, Dec 13, 2024

Enjoyed what you've read so far? Great news - there's more to explore!

Stay up to date with the latest news, a vast collection of tech articles including introductory guides, product reviews, trends and more, thought-provoking interviews, hottest AI blogs and entertaining tech memes.

Plus, get access to branded insights such as informative white papers, intriguing case studies, in-depth reports, enlightening videos and exciting events and webinars from industry-leading global brands.

Dive into TechDogs' treasure trove today and Know Your World of technology!

Disclaimer - Reference to any specific product, software or entity does not constitute an endorsement or recommendation by TechDogs nor should any data or content published be relied upon. The views expressed by TechDogs' members and guests are their own and their appearance on our site does not imply an endorsement of them or any entity they represent. Views and opinions expressed by TechDogs' Authors are those of the Authors and do not necessarily reflect the view of TechDogs or any of its officials. While we aim to provide valuable and helpful information, some content on TechDogs' site may not have been thoroughly reviewed for every detail or aspect. We encourage users to verify any information independently where necessary.

Tags:

Emerging Technology Artificial Intelligence AI Generative Artificial Intelligence Generative AI GenAI Meta FAIR Meta Fundamental AI Research Research Code Models Datasets Motivo Audio Seal Movie Gen Audiobox Melody Flow Explore Theory-of-Mind Large Language Model LLM Large Concept Model LCM Dynamic Byte Latent Transformer Memory Layers Image Diversity Modeling CLIP 1.2 Metaverse

Join The Discussion

Artificial Intelligence

FeaturedArtificial Intelligence

Artificial Intelligence

FeaturedArtificial Intelligence

Artificial Intelligence