Meta AI Releases Data2vec 2.0: An Efficient Self-Supervised Learning For Machine Learning Tasks

Meta AI recently introduced Data2vec 2.0, an AI processor that uses computer vision for multi-camera multi-person re-identification. Data2vec 2.0 is 16 times faster than the current leading self-supervised method in computer vision and is just as accurate. It is 11 times faster than wav2vec 2.0 on the LibriSpeech benchmark and just as accurate as RoBERTa on the GLUE benchmark while requiring less training time. Data2vec 2.0’s code and pretrained models are open-sourced, with the hope that it will help machines understand complicated data better.




Étiquettes :