One of the most frustrating thing for an online viewer is having to wait for the video to buffer. This happens because videos are loaded into small chunks that have been broken down by special algorithms. With a slow internet connection, you may experience a few moments of lower resolution so the video can keep running. Skipping ahead at this point won’t work as the video will need to stall until it’s completed buffering that part.


YouTube is just one example of an internet business that makes use of these adaptive bitrates (ABR) algorithms and it does this to try and give its customers the best possible viewing experience. It’s also a great way to save bandwidth too. This is because most people don’t actually watch videos right to the end, so it wouldn’t make any sense to buffer the whole thing straight away every time, would it? But, companies still need to be careful not to put users off by unreasonable amounts of buffering.

“Studies show that users abandon video sessions if the quality is too low, leading to major losses in ad revenue for content providers,” says MIT Professor Mohammad Alizadeh.  So, he set out to solve this issue along with colleagues at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). Together they developed “Pensieve”, a machine learning artificial intelligence (AI) system that has the ability to choose different algorithms depending on the conditions of the network at that time.  With this system in place, users experience much less rebuffering than with previous systems.


Another great feature of Pensieve is that you can also customize it based on the content provider’s priorities. If for example, you were about to go through a tunnel where signals are weak, YouTube could turn down the bitrate so enough of the video can be loaded and there will be no need to rebuffer during the temporary loss of network. “You could even imagine a user personalizing their own streaming experience based on whether they want to prioritize rebuffering versus resolution,” says Hongzi Mao, Ph.D. student and lead author on a related paper with Alizadeh and Ph.D. student Ravi Netravali.

In a nutshell, ABR algorithms come in two types: ones that are rate based and measure how fast data is transmitted across a network, and buffer-based ones that make sure there’s always a certain amount of future video that’s been buffered already. However, these are both limited in what they do and often require human experts to enable the network to adapt to different conditions. Pensieve, on the other hand, doesn’t suffer from these same issues. It’s ABR algorithm is presented as a neural network that continually tests out various buffering and network speed conditions. Mao, who was the lead author on the paper, explained, “It learns how different strategies impact performance, and, by looking at actual past performance, it can improve its decision-making policies in a much more robust way.”


The team behind Pensieve have been testing it extensively in various settings including using an LTE network while outside and using WiFi at a local cafe and the results are that it can achieve the same resolution as MPC, but with up to 30 percent less rebuffering. “This work shows the early promise of a machine-learned approach that leverages new ‘deep learning’-like techniques.” It also demonstrates how Pensieve will operate effectively in situations that are new to it. Moving forward the team will be looking to see how Pensieve behaves when tested on virtual reality (VR) video.

More News to Read

Comments

comments