Streaming video Online takes up lots of expensive computing & network resources, & a user’s experience can be really poor because of a video being pixelated or due to buffering. Now MIT has developed ‘Pensieve’, an Artificial Intelligence (AI) system run on machine learning that picks the optimal algorithms based on network conditions. A higher quality streaming experience is now possible with less rebuffering than existing systems.
The algorithms employed by popular streaming services like YouTube or Netflix break videos up into small chunks which get loaded up as you watch. When your Net connection is slow, your resolution may get reduced to make sure you still get an uninterrupted viewing experience, which is why you’ll experience pixelation. When you skip ahead further into the video you start a rebuffering process & you may suffer from pixelation again.
These adaptive bitrate (ABR) algorithms save bandwidth & give users a consistent viewing experience, while saving bandwidth. The trade-off is that perceived video quality can often be low, resulting in lost revenue when viewers prematurely abandon clips, so researchers have been looking for better, cheaper alternatives.
To counter all this, MIT’s Computer Science & Artificial Intelligence Laboratory (CSAIL) led by Professor Mohammad Alizadeh has come up with Pensieve. The system can stream video with 10-30% less rebuffering than current levels, with levels rated 10-25% higher by users on ‘quality of experience’ metrics.
Pensieve doesn’t just rely on the currently popular model predictive control (MPC) or ABR algorithms, which rely on static models, opting instead for a neural network approach. According to 1 of the other lead researchers at MIT, Hongzi Mao, “It learns how different strategies impact performance, &, by looking at actual past performance, it can improve its decision-making policies in a much more robust way”.
Instead of relying on the intuition of human experts, Pensieve uses “a machine-learned approach that leverages new ‘deep learning’-like techniques”.
It tunes its algorithms employing rewards & penalties, so that it will get rewards for high resolution & buffer-free viewings & penalties when it rebuffers. A streaming provider can fine tune the reward system based on the metrics they think users will appreciate. The example given by the researchers is, “…studies show that viewers are more accepting of rebuffering early in the video than later, so the algorithm could be tweaked to give a larger penalty for rebuffering over time.”
Check out this video:
This approach will soon make its way to YouTube & Netflix & other streaming services, but meanwhile the MIT researchers want to apply it to high resolution VR (4K). Current network bandwidth doesn’t support raw 4K streams & the researchers are hoping that their approach can kick-start immersive online VR experiences.