In the process of seeking human intelligent AI, meta bet on self supervised learning

Meta AI Mask self encoder for computer vision is trained on most occluded images [ Left ]. However , Its reconstruction [ center ] Very close to the original image [ Right ].

Meta Chief AI scientists Yann LeCun Once expressed such a vision , “ We want to build intelligent machines that learn like animals and humans ”. In order to achieve this vision ,Meta Put the treasure on Self supervised learning （SSL） On .

Some time ago , A title is MAE This paper shows how the self-monitoring system can reconstruct images from very scattered and incomplete data . However ,MAE In fact, it is not a new concept ,Meta This work has already been extended to new fields .

In video, for example MAE in , The mask rate of each video frame is as high as 95%, Because the similarity between frames means that video signals have more redundancy than static images . and MAE By shielding up to 95% Each frame , Reduce computing costs by up to 95%.

Another example is in audio MAE in ,Meta AI The team converts sound files into spectrograms , Partially mask it for training . The reconstructed audio is impressive , Although the model can only process clips for a few seconds . The team said it would soon be in arxiv.org Publish audio on MAE Related work of .

 

Link to the original text ：https://spectrum.ieee.org/unsupervised-learning-meta