Meet MultiRay, Meta’s new platform for efficiently running large-scale, state-of-the-art AI models. By converting input to an embedding that’s cached & shared between users, MultiRay greatly reduces complexity & energy use. Here’s how it's enabling us to build & scale ML work
@aiatmeta
-
ImageNetX: Explaining Vision Model Failures Through Detailed Annotations
By
–
Even today’s best #deeplearning vision systems can fail when pose/lighting/background vary. Existing benchmarks show challenging examples but don’t explain why. ImageNetX includes annotations for the entire ImageNet1k validation set + a random subset of 12k training images.
-
ImageNetX: Human Annotations for AI Vision Model Robustness Analysis
By
–
We’ve released ImageNetX: a set of human annotations for the popular ImageNet benchmark to gauge model robustness strengths/weaknesses — one of the first large scale efforts to pinpoint mistake types in AI computer vision systems. Explore the dataset
-
NLLB-200 Achieves Lowest Content Deletion Rate Among Translation Services
By
–
NLLB-200 sees only 0.13% of translated content deleted. That’s the lowest percentage across all machine translation services available on the platform, suggesting that the resulting translations are being understood & accepted. 3/5
-
NLLB-200 Achieves Superior Translation Quality Across All Languages
By
–
Across all languages, NLLB-200 is seeing the best results for translations modified <10% compared to all other MT services on the platform — a strong signal for the quality of translations that are being generated. 4/5
-
NLLB-200 becomes third most-used translation engine in four months
By
–
NLLB-200 now represents 3.8% of all machine translations on the platform. This makes it the third most-used machine translation engine across all published translations just four months after launch. 2/5
-
No Language Left Behind Improves Wikipedia Translation Tool Usage
By
–
In July, we launched No Language Left Behind to more languages in @Wikimedia
’s Content Translation tool that helps Wikipedia editors jumpstart article translation — a new report is already showing encouraging impact on usage & translation quality https://
bit.ly/3NO63VR /5 -
Meta AI Creates Largest Protein Language Model with 15B Parameters
By
–
Meta AI researchers trained a language model to fill in protein sequence gaps across millions of diverse proteins & scaled up to 15B parameters, creating the largest language model of proteins to date. More on our latest breakthrough in protein folding https://
bit.ly/3WoWcK2 -
Team Receives Honorable Mention at ECCV2022 for Pose-NDF
By
–
Proud of the team and the honorable mention they received last week at #ECCV2022 for their great work on Pose-NDF! You can learn more about the model and read the paper at the link in Garvita's tweet
-
Model Available in Lean VSCode Plugin for Researchers
By
–
We’ve made our model available through the Lean VSCode plugin to allow researchers to explore the capabilities within the Lean environment. We hope you’ll build on this work as part of our shared pursuit of rapid progress in this exciting field.