Data2vec 2.0 can train self-supervised speech, vision & text models up to 16x faster than the most popular existing algorithm for images — achieving the same accuracy. Read more & get the open source code https://
bit.ly/3H9mPf7
@aiatmeta
-
Data2vec 2.0: 16x Faster Self-Supervised Learning Across Modalities
By
–
-
E3B: New SOTA Method for Complex Environment Exploration
By
–
E3B is a method for exploring complex environments which vary across episodes. This work set a new SOTA for MiniHack & reward-free exploration on Habitat in October.
— AI at Meta (@AIatMeta) 23 janvier 2023
Read the paper ➡️ https://t.co/hLIEDsoTwl
See the code ➡️ https://t.co/vVR3Kw6AQL pic.twitter.com/KPREx5srALE3B is a method for exploring complex environments which vary across episodes. This work set a new SOTA for MiniHack & reward-free exploration on Habitat in October.
Read the paper https://
arxiv.org/abs/2210.05805
See the code https://
github.com/facebookresear
ch/e3b
… -
Meta AI Residency Program Applications Open Until January 24th
By
–
Do you know someone who's interested in… Pursuing a PhD? Publishing research? Getting hands-on experience with AI research? Applications for the one-year Meta AI Residency close on January 24th! Learn more and apply here
-
ESM Metagenomic Atlas: Mapping the Protein Universe’s Dark Matter
By
–
The ESM Metagenomic Atlas is the first comprehensive view of the ‘dark matter’ of the protein universe. Made possible by ESMFold, a breakthrough model for protein folding by Meta AI that allowed us to accelerate folding by up to 60x.
— AI at Meta (@AIatMeta) 19 janvier 2023
More on this work ➡️ https://t.co/v8CRwWXfpY pic.twitter.com/WtAfGL0gDNThe ESM Metagenomic Atlas is the first comprehensive view of the ‘dark matter’ of the protein universe. Made possible by ESMFold, a breakthrough model for protein folding by Meta AI that allowed us to accelerate folding by up to 60x. More on this work https://
bit.ly/3ZRCVmu -
AI-Powered Audio Hypercompression Achieves 10x Better Rate Than MP3
By
–
ICYMI, we shared some of our progress on AI-powered hypercompression of audio. We achieved ~10x compression rate vs. MP3 at 64 kbps with no quality loss — the first time this had been done with 48 kHz sampled stereo audio.
— AI at Meta (@AIatMeta) 13 janvier 2023
More on our blog ➡️ https://t.co/wEncd6rAbP pic.twitter.com/uBLfHMsD1bICYMI, we shared some of our progress on AI-powered hypercompression of audio. We achieved ~10x compression rate vs. MP3 at 64 kbps with no quality loss — the first time this had been done with 48 kHz sampled stereo audio. More on our blog https://
bit.ly/3FfgPCb -
Meta AI Presents CICERO: First Human-Level Diplomacy Game AI
By
–
Meta AI presents CICERO — the first AI to achieve human-level performance in Diplomacy, a strategy game which requires building trust, negotiating and cooperating with multiple players. Learn more about #CICERObyMetaAI: http://
bit.ly/3GBwLzx -
Meta’s Variance Reduction System for Equitable Ad Distribution
By
–
Today, we’re sharing more about the Variance Reduction System — a new offline RL framework developed by Meta to help ensure a more equitable distribution of ads on our services. More on how it works, the challenges yet to be solved & why AI fairness is such an important field
-
AI Solves 10 International Math Olympiad Problems with HyperTree Method
By
–
Using our HyperTree Proof Search method we created a neural theorem solver that was able to solve 10 International Math Olympiad problems — 5x any previous AI system. More in this post ➡️ https://t.co/dec9VeJhny pic.twitter.com/rns9IffWqx
— AI at Meta (@AIatMeta) 3 janvier 2023Using our HyperTree Proof Search method we created a neural theorem solver that was able to solve 10 International Math Olympiad problems — 5x any previous AI system. More in this post https://
bit.ly/3vBnu3E -
Thank You for Supporting AI Research and Innovation in 2022
By
–
From all of us to the entire community: Thank you for supporting AI research, innovation & open science in 2022 — we're excited to continue pushing this (and more) forward together in 2023.
-
OPT-IML: Instruction-Tuned Models for Diverse Language Tasks
By
–
For OPT-IML, we boosted performance of our OPT-175B work using instruction tuning, allowing us to adapt it for more diverse language applications (Ex. Answering Q’s, summarizing text & translation). A 30B parameter model is available today with a 175B model available soon.