Buffering is a technical annoyance we've learned to tolerate, but deep down we hate it with fervor. No matter how fast the connection, software, or device, video streaming always finds a way to "choke" and ruin the experience. Major services apply adaptive bitrate mechanisms to compensate, but at MIT they have a broader vision. The project is called Pensieve, and its goal is to implement optimal dynamic bitrate algorithms with the help of a neural network.
The Streaming Lottery
Streaming on YouTube is a real lottery for me. Most of the time it picks 480p when the connection has no trouble handling a 720p or 1080p stream, but it has offered 240p for no reason. Unfortunately, things don't improve much with its buffering, and we all agree. The performance is too erratic, and as for pauses, well... they test our patience. The point is we demand more from streaming. Better quality, greater fluidity, and superior intelligence when implementing dynamic adjustments. Honestly, I think we'd settle for much more efficient and consistent buffering. That's where MIT and its recent Pensieve project come in.
How Pensieve Works
Technically, Pensieve is another adaptive bitrate system for video streaming, but the difference is that it bases its decisions on "machine learning". According to the study, Pensieve trains a neural network model that chooses the bitrates for future video "chunks", using observations recorded through players and the performance from previous selections. Pensieve's accuracy depends on the amount of material it's exposed to. The researchers prepared their system with the equivalent of a month of video, and they've already achieved a reduction in buffering that ranges from 10 to 30 percent.
The Road Ahead
The researchers will present their full work at the SIGCOMM conference on August 21, and it seems they'll make Pensieve open source. That should significantly expand its flexibility, enabling additional optimizations.
