//
What kind of innovation has the explosion of ChatGPT brought to audio and video? How will AIGC assist audio and video technology to bring users a better experience? On July 29th, LiveVideoStackCon Shanghai Station AI and Multimedia Session will answer your questions.
AI and multimedia
With changes in lifestyle and habits, people have higher demands and expectations for media content and experience. How to provide higher video quality and immersive audio and video experience has become a challenge for the development of the new generation of multimedia. Artificial intelligence, computer The development and breakthrough of technologies such as vision and cloud computing provide more opportunities and possibilities for these challenges.
Topic introduction
TOPIC1 "Family Scene Large Model Technology and Application Practice"
Fu Tao, Artificial Intelligence Expert of China Mobile Smart Home Operation Center
As a new way of life, smart home has received more and more attention.
Modules such as dialogue and companionship, family security, family education, family health and family entertainment are important components of smart homes, which involve multi-modal and complex AI understanding and generation technologies such as natural language processing, computer vision, and speech processing . The continuous improvement of intelligence requirements requires the use of general intelligence emerging from large-scale pre-trained models, and targeted fine-tuning and adaptation of scenarios; specifically, we use large models with hundreds of billions of parameters and combine them with family scenarios. Instruction fine-tuning and corresponding algorithm optimization have empowered hundreds of millions of smart home users, among which the coverage of smart dialogue, smart security, and family education has reached more than tens of millions;
This sharing is divided into 4 parts. The first part mainly introduces the core business of the family scene and simultaneously introduces the key issues of algorithm application. The second part mainly reviews the development status of the industry's large model and the main algorithm architecture. The third part is our In the research and development of key technologies of large models in the family scene, the fourth part is mainly about the application practice of large models.
TOPIC2 "Practice of Optimizing Image Quality and Bandwidth Based on Device-Cloud Combination Based on Perceived Quality of Human Eyes"
Jianhan Xiaohongshu Head of Audio and Video Architecture Video Image Processing Algorithm
With the large-scale growth of Xiaohongshu's video business and short video playback, how to effectively improve user experience quality while reducing video bandwidth costs has become an important technical optimization goal. On-device super-resolution is an effective technology for improving image quality and saving bandwidth. However, computing resources and performance constraints on the device limit the implementation of advanced super-resolution algorithms. In addition, how to evaluate the video quality after super-resolution on a large scale so as to guarantee and quantify the quality of user experience to the greatest extent has also become a problem to be solved.
To meet the above challenges, Xiaohongshu’s audio and video architecture algorithm team has developed a device-cloud combined super-resolution framework based on human perception quality , which effectively compensates and improves the device-side super-resolution effect through cloud video processing, and uses AI-based no-reference video quality The evaluation index achieves quality evaluation consistent with human vision, providing key support for cloud image quality analysis and processing and large-scale quantification of user experience quality. The overall solution achieves quantifiable end-side super-resolution effects, high integrated coverage, and high bandwidth savings.
TOPIC3 "Data is Technology - AIGC Bottom Data Exploration"
Qingqing ZhangFounder and CEO of Magic Data
After the outbreak of ChatGPT, there was an upsurge of large models around the world. A number of technology giants including Microsoft, Google, Baidu, and Ali have accelerated the refining of "big models". Data is one of the three elements of large-scale model construction and a key factor determining model intelligence. At present, there are still many challenges in domestic model training data, such as scarcity of corpus, low data quality and data compliance.
This sharing will be mainly divided into four parts. First, it will introduce the current development status of the large model and the pain points of the current training data. Then it will introduce the comparison between the domestic Chinese data and the data used by ChatGPT. The third part will introduce the conversational data through experiments. The performance ability in large model training, and finally introduces how to build data-centric MLOps to help large model construction and landing.
TOPIC4 "Smart Outbound Calls, Leading the Future of Credit Services"
Chen Liang Dongting Intelligent Algorithm Engineer
With the development of artificial intelligence technology, the model of enterprise customer service is also changing. The traditional customer service model is often manual agents answering calls or online customer service replying messages, but this model is inefficient and prone to human errors, and does not meet the needs of modern enterprises for efficient and accurate customer service.
This sharing will be divided into three parts. The first part introduces the pain points of the traditional customer service model and how the XiaoGo smart outbound robot solves these problems. The second part introduces the advantages and disadvantages of the XiaoGo compared to the traditional customer service scenario Highlights, the third part introduces technical features. Through the sharing of the above three parts, we will give you a comprehensive introduction to the innovative application and technical advantages of the XiaoGo intelligent outbound robot in traditional customer service scenarios.
▲Scan the QR code in the picture or click " Read the original text " ▲
sign up now