AI-based Image and Video Compression on Mobile Neural Accelerators

点击翻译按钮获取中文摘要

Deep Render is a research-led startup founded by Imperial College graduates from the Department of Computing. We are a London based AI start-up that is developing the next generation of media compression algorithms. Our proprietary and patented technology is at the forefront of machine learning research, and our Biological Compression has already leapfrogged the best previous compression standards by up to 80% better efficiency. Additionally, while the traditional compression methods have hit peak-innovation with significantly declining performance gains over the past decades, Biological Compression is only starting to realise its full potential.As global data consumption is growing exponentially with more than four out of five bits of global internet traffic being image and video related, Deep Render's AI-based compression technology is vital to counteract upcoming broadband constraints. The outbreak of COVID-19 has accelerated this problematic trend even further, as a result of the crisis, internet usage has increased significantly. In particular, the demand for streaming, video-on-demand and remote-working tools has skyrocketed. The exponential increasing bandwidth demand creates a massive challenge, as seen by the EU's request asking big-tech companies to lower streaming resolutions to safeguard our communications infrastructure. Better compression technology is urgently needed to stop this problem from snowballing. Jeff Hecht states in Nature.com "Researchers are scrambling to repair and expand data pipes worldwide to keep the information revolution from grinding to a halt".This project is about implementing Deep Render's prior compression research on consumer products so that the average person can benefit from smaller file sizes, and consequently, higher bandwidth supply. Precisely, we aim to execute our software on neural accelerator chipsets of an iPhone 12 and a Galaxy S20\. The challenge is the performance-optimisation of our computational-hungry algorithm (neural networks) and to make it runnable in real-time on mobile platforms without loss of our superior compression efficiency. This task requires in-depth knowledge of mobile chipsets, as well as low-level software optimisation and engineering capability.To accomplish our goal, we propose a collaboration with Professor Paul Kelly from Imperial College London. Paul Kelly is a Professor of Software Technology, he is the lead for the Software Performance Optimisation Group, and he has a close affiliation to Imperial's Artificial Intelligence Network. We aim at leveraging his extraordinary expertise in high-performance software implementation and his prior research in low-level neural network optimisation to reach sub-16.6ms decoding times for 4K video on and iPhone 12 and Galaxy S20 device. Professor Kelly sits on the Advisory Board of Deep Render since June 2018 and has an excellent understanding of Deep Render's technology, and a great working relationship with its founders.Our value proposition is easy to understand. By making file sizes 80% smaller, we increase the bandwidth supply of the internet by a factor of up to 5\. Deep Render is going to help create a new age in which bandwidth constraints are a problem of the past. As a result of COVID-19, solving this problem has gained more importance, and Deep Render is determined to create a fast solution.