MIT develops a new AI – a neural adaptive system, dubbed “Pensieve”, to stream high-quality video online with less rebuffering. Earlier experiments have shown impressive results – 10 to 30% decrease in rebuffering of video streams.
Not everyone has low-latency, high speed internet connection and this usually results in a lot of lag and pixelated video, which is extremely irritating.
To diminish this bad experience, entire video is not sent to your device at once. It would take too much bandwidth. People don’t watch every video till the end. On big sites like YouTube, about 1 billion hours of video content are streamed daily, it would be a huge waste of resources to buffer millions of long videos for all viewers every time. So instead, video data is sliced into smaller pieces and delivered sequentially.
The Current Method
To ensure you get the best quality without any buffering, platforms like YouTube leverages Adaptive BitRate (ABR) algorithms. These types of algorithms can be divided into 2 categories-
- Calculates how fast your Internet connection can transfer data, and
- Maintain enough buffer at the head of the video.
The video suffers pixelation when rate-based algorithm fails. It drops the bitrate (quality) to make sure that the video is still playing. When you skip to any particular part of the video, it causes complete destruction of buffer-based algorithms and they have to quick-freeze playback while they gather both the new smaller pieces of video and the buffer ahead of it. This is the main problem with the current algorithms and neither of them is completely capable of resolving it. This is the part where artificial intelligence comes in.
Other Not-So-Effective Models
There’s already been a lot of research and experiments to address this issue. One of them is Model Predictive Control (MPC). It predicts the network conditions – how network speed will change over time and optimize the decisions based on the predictions. However, the system is not suitable for networks that experience sudden traffic change or intense fluctuation of data flow.
Pensieve (developed by Computer Science and Artificial Intelligence Lab, MIT) doesn’t rely on any model. Instead, it uses machine learning to determine when (and under which circumstances) to switch between buffer-based and rate-based ABRs.
Just like other neural network, Pensieve uses penalties and rewards to weight the results at each outcome. The system changes its behavior to receive the highest reward at the end. Because the rewards can be altered, the whole system can be changed to behave in any manner we want.
For instance, the system might get a reward when it delivers a high-resolution, buffer-free video, but a penalty if it has to rebuffer.
Developers have trained this artificial intelligence on ‘a month’s worth of downloaded video’. They have tested it under various scenarios, including LTE network while walking down the street and WiFi at a cafe. It returned the same video quality as MPC system, but with 10% to 30% less buffering.
‘This new system is flexible to adapt anything a user wants to optimize it for. A viewer can even personalize his streaming experience by prioritizing resolution vs rebuffering’, said Mohammad Alizadeh, MIT professor.
‘The system learns how various strategies affect performance, and by analyzing the past performance, it can improvise the process of decision-making in a more robust manner’, said Mao, lead author of the new study.
What Developers are Looking for?
It is expected that Pensieve will be adopted by sites like Netflix and YouTube, but MIT team is hoping to apply the artificial intelligence to virtual reality. The bit rates require for playing 4K quality VR cross 100 Mbps, which current networks can’t simply support. ‘The team is excited to see what Pensieve can do for technologies like VR. This is only the first step in seeing what smart neural adaptive systems can do’, said Alizadeh.