In a recent development with YouTube, DeepMind’s MuZero AI has made significant progress towards mastering video compression. This could potentially revolutionize the way videos are stored and streamed over the internet.
Video streaming has become a massive part of internet traffic, with analysts predicting that in 2021, streaming video would make up the majority of internet traffic. With the continued growth of internet traffic, the need for efficient video compression becomes even more important.
MuZero’s collaboration with YouTube has already resulted in a 4% bitrate reduction across a wide range of videos. This kind of reduction has numerous benefits, including faster loading times, better resolution, and reduced data usage.
Most videos on the internet rely on codecs to compress and encode the video for transmission and then decode it for playback. MuZero’s application of reinforcement learning to optimize video compression is a promising development for the future of efficient video streaming.
The initial focus of MuZero’s work is on the VP9 codec, which is widely used by YouTube and other streaming services. This is a significant step forward in the application of AI to real-world problems beyond just solving games.
The algorithm used by MuZero for video compression is based on the rate control module within the VP9 codec. This module works to optimize bitrate through a parameter called the Quantisation Parameter (QP).
MuZero’s ability to combine search power and the learning of a model of the environment make it an ideal candidate for solving the issue of rate control in video compression.
One of the challenges in applying MuZero to video compression lies in the variability of uploaded videos, each with different content and quality. DeepMind has developed a mechanism called self-competition to address this challenge, simplifying the complex objective of video compression into a simple WIN/LOSS signal.
Overall, MuZero’s progress in optimizing video compression is just the first step towards solving real-world problems with the help of AI. DeepMind envisions a future where a single algorithm can improve a wide range of real-world systems across various domains.