Kuaishou Delivers a Keynote Speech at the "The Fire of Innovation" Held by GeekPark About Artificial Intelligence at the Core of Short-Video Platforms

MOUNTAIN VIEW, Calif., Oct. 9, 2018 /PRNewswire/ -- On September 18, Wen Zheng, Kuaishou's Vice President of AI Technology, delivered a keynote speech at "The Fire of Innovation" forum held by GeekPark where he asked the question, "How does AI truly empower short video platforms?" Kuaishou, also known as Kwai, started out seven years ago and has evolved from just a tool app for making and sharing GIFs to a massive short-video community with more than 100 million daily active users, each spending more than 60 minutes daily on the platform.

In seven years, Kuaishou users have already uploaded more than 7 billion short videos with content ranging from short comedic scenes, games, and entertainment to local customs and people. It is not a surprise that many people tend to think of Kuaishou as just a short video platform that has little to do with AI technology.

However, AI technology is actually essential to effectively allocate each user's time and attention to the vast amount of videos instead of just a small number of popular ones. This would have been impossible with mere human operations.

AI Adds More Fun to Capturing Life Moments

With tens of millions of new videos being uploaded every day, it is an immense challenge to match the right videos to the right users. To address this, Kuaishou has designed a complete and all-rounded set of AI-based solutions for all its operations covering video production, content comprehension, user understanding, and system distribution. AI is the core capability that connects content generation and content consumption by the user, said Wen Zheng at the Geek Employer Exchange Meeting.

Kuaishou also recently developed and launched a series of highly popular special video effects. One of them is called the "Kwai Time Machine" creates an "aging" effect that transforms characters in the video into what they would look like 60 years later in just a dozen seconds. Other special video effects include dancing games that support real-time recognition of body motion, as well as AR face changing effects. These features reflect Kuaishou's capabilities in developing and introducing cutting-edge AI technologies that involve complex and multiple technical modules such as human pose estimation, hand gesture recognition and background segmentation. With these innovations in content production, Kuaishou is trying to spice up to the video-capturing process.

When a short video is created and uploaded onto the app, the back-end system will extract basic information about the video, such as the gender of the people captured and their expressions in order to try to understand the video content. The machine will also further classify the videos into various categories via scenario recognition, object tracking, and image quality assessment. Another important aspect that will help machine understand the video is speech recognition. The machine is actually able to convert dialogue in the video into text to understand what the video is about. Just like how human beings save whatever they learn in their brain, Kuaishou learns and organizes the myriad of content on the platform and stores them in its knowledge graph. By integrating perceptual content with the knowledge graph, high-level semantics and sentiments within the videos can be identified.

It is also necessary for the machine to understand users. The huge volume of behavioral data generated by users will be fed into a deep learning model for machine training, so as to gain comprehensive user information to predict users' preferences and the connection between individual users.

Kwai Breaks the "Information Cocoon"

From making video-recording more interesting, to enabling the machine to understand video content and gain in-depth visibility into users, these are Kwai's attempts at empowering its short video platform with AI technology.

According to Wen Zheng, an efficient process has yet to be established for the development and implementation of AI, which is still a new technology. As for itself, Kwai carries out problem-oriented AI projects, which starts with identifying requirements, followed by requirement-based pre-research, data collection and model training, and then the development and productization of corresponding capabilities.

Artificial intelligence and algorithm recommendation have aroused some controversy. It is generally believed that, as intelligent algorithm recommends content based on users' interests, over time, users will be surrounded by things they care about and lack an overall understanding of the outside world, leading to the formation of an "information cocoon". In response, Wen Zheng introduced their recommendation mechanism which goes beyond providing content that interests users the most to considering the diversity of content and exploring things of interest in a broader range.

For example, a user may find some other users in the community who are similar to them, and then become interested in what they like. Therefore, the user is not alone in the discovery journey, but with thousands of similar yet different people. That's why a distribution system based on this logic can really take into account the long tail videos, allowing more people to get public attention.

As a record carrier, the short video itself is inclusive. It dramatically lowers the threshold for information exchange and empowers people to record and share their life moments in a fair and interesting way using technology.

This April, Kwai and Tsinghua University jointly established the Institute of Future Media Data. To conquer the technical challenges faced by School of Software, Tsinghua University and Kwai, the institute will collaborate with the School of Journalism and the Department of Sociology to conduct research in the fields of society and humanities, covering multiple areas such as computer graphics, image processing, big data and artificial intelligence. It aims to explore more possibilities for technology-based video-streaming in a cooperative and open manner.

SOURCE GeekPark