Play a variety of application scenarios, Kuaishou MMU shares secrets

✏️ Editor's note:

The rise of short videos has witnessed an information revolution. The era of graphics and text is gradually transitioning to the era of multimedia, and understanding and searching for video has become a key technology today. As the AI middle-end of the national-level short video app, how does the Kuaishou MMU (Multimedia understanding) team cope with various application scenarios? The Milvus community is honored to invite Yu Jin, a research and development engineer from Kuaishou MMU, to share with you the application of Milvus in Kuaishou.

🌟 Guest profile: Yu Jin, Kuaishou multimedia content understanding engineer, responsible for engine architecture and large-scale vector computing, graduated from Peking University, likes reading and jogging.

For the full video, please click: https://www.bilibili.com/video/BV1444y1p7c8?spm_id_from=333.999.0.0

MMUMMR1.0: Vector approximation computing platform based on Milvus database

Kuaishou MMU is an AI middle-end platform responsible for the Kuaishou short video search system and video understanding. Its business covers basic AI algorithms such as OCR, ASR, word segmentation, and NER; middle-end technologies such as short video classification and labeling system construction; and system services such as short video search. . In this AI middle-end, vector computing plays a crucial role.

The Kuaishou MMU team needs to deal with many application scenarios related to vector computing: similar video retrieval, video compliance retrieval, original video detection, commodity detection... Before contacting Milvus, the team used a self-developed vector retrieval system, but the implementation methods were compared Complex, high maintenance cost, and general system availability, a high-performance, easy-to-access and high-stability vector database is urgently needed for use by various business parties.

After a series of product researches, the Kuaishou MMU team finally chose Milvus, which has an active community, stability and performance, as the AI middle-end platform to build platforms including AI models, data analysis tools, and ANNS. The scenarios that have been implemented so far include hundreds of billions of video retrievals and billions of commodity retrievals, and more scenarios will gradually migrate to the ANNS platform built by Milvus.

The MMUMMR 1.0 architecture based on Milvus 1.1 is shown in the figure above, in which the vector data storage calculation is based on Milvus's data fragmentation and merging management. The Milvus database supports a cloud-native distributed architecture, with the characteristics of separation of storage and computing, separation of writing, construction, distributed computing, and query, integration of streams and batches, and elastic scaling.
(Milvus 2.0 also met with you not long ago, let's take a look at the new features of Milvus 2.0 !)

Champion Scenario Analysis: What Are We Talking About When We Talk About Vector Computation

The Kuaishou team won the first place on the track in the first international vector retrieval competition held not long ago. The track requires teams to target at least three of the six billion-scale datasets to achieve performance above 10,000 QPS while improving the recall rate as much as possible relative to the benchmark scheme Faiss' IVFPQ method. The Kuaishou team's solution is fully optimized for the IVFPQ method, and is 5% - 10% higher than the baseline on all four datasets.

For more information about the first international vector retrieval competition, please refer to: -University-Research Cross-border Dialogue, Review of Vector Database Symposium

Introduction and Application of Vector Approximate Computing API

The Milvus database provides a complete set of simple and intuitive APIs. The Collection index library of MMUMMR 1.0 uses Milvus native APIs such as create, drop, count, and stat to realize automatic lifecycle management business applications such as feature recall of cold-start videos in the last three days and feature recall of recent popular videos in the last 90 days.
The Milvus Collection API address is: https://milvus.io/api-reference/node/v2.0.0/api%20reference/collection.html

In addition, MMUMMR 1.0 also adopts high-precision KNN retrieval in video duplication check, video generation material retrieval, commodity SKU identification, parallel corpus construction and other businesses; uses attribute retrieval in video commodity retrieval and video retrieval scenarios; deduplicates video covers The ADBKmeans clustering method is used in the business. For details, please refer to the video👇

https://www.bilibili.com/video/BV1444y1p7c8?spm_id_from=333.999.0.0