Visit Baidu

Super slow training is the bottleneck in meta-learning. We introduce Faster Meta Update Strategy (FaMUS) to replace the most expensive step in the meta gradient computation with a faster layer-wise approximation. Read our #CVPR2021 paper for more: https://t.co/pOkt5D36Ji https://t.co/HDNb1SmDJ9

Introducing PP-YOLOv2, our upgraded PP-YOLO object detector. The v2 boosts performance from 45.9% - 49.5% mAP, runs on 68.9FPS at 640x640 input size & achieves better balance between speed & accuracy than YOLOv4 & v5. Paper: https://t.co/CaHi672osP Code: https://t.co/dlPPSiV5qv https://t.co/9Zwq1fPUh9

We've released BSTC, a large-scale ZH-EN speech translation dataset. It includes 68 hours of Mandarin speech from licensed videos, the videos' human-annotated transcripts and their translations, as well as their ASR results. Read the paper: https://t.co/KahK2g5aG7 https://t.co/8qdgZyi9Xl

We release Diffwave, a SOTA audio synthesizer that is 1000x faster than WaveNet and 20x smaller than WaveGlow. It also outperforms GAN-based waveform models in more challenging unconditional generation tasks. Read our #ICLR2021 paper: https://t.co/Og6RibIjo5 https://t.co/6r4g1yBTjX

We've discovered some interesting phenomena in contextualized word embedding (BERTs, ERNIEs, GPTs, etc.) spaces. They form clusters near extremely low dimensional manifolds and are more isotropic than people thought previously. Read our #ICLR2021 paper: https://t.co/WyXK1v0L1P https://t.co/XVYpTy0kGW

In a recent paper, we tackle the problem of semi-supervised transfer learning with a practical & competitive paradigm that utilizes pre-trained models in semi-supervised learning. Our method outperforms SOTA SSL techniques in real-world tasks. #cvpr21 https://t.co/4qL3D6ibkl https://t.co/8mrLisVEnc