Algorithm Powers Innovation

To solve the new optimization problems on the cloud, including online resource provisioning, offline resource sorting, etc.., we need to optimize the existing operations optimization algorithms or use reinforcement learning algorithms popular in recent years based on
data-driven solutions.

Using machine learning to profile users, VMs, storage and other
resources to make resource allocation more efficient.

Researching on how to use machine learning to detect and recognize sound events, such as baby crying, glass breaking, animal calls, etc. Some of the algorithms has already been successfully applied to Tech4all rainforest projects (detecting the sound of chainsaws, spider monkey calls).

Domain custom ASR engine consists of flow recognition, phrase sound recognition and long speech recognition, which are widely used in customer service agents, finance and other fields.

The user-defined keyword spotting algorithm can generate a customized model without training corpus and can be deployed on the edge and cloud side as needed to meet the user's personalized needs.

Exploring the architecture and algorithm of a new generation of learned system with self-learning, self-adaption and self-optimization capabilities for the cloud computing system. Promoting it from heuristic architecture and algorithm to learning architecture and algorithm. The algorithm has been successfully applied in learned partition, learned sorting and construction of learned multidimensional index. Cloud data services
are being further explored and researched now.

CDN will become a media intelligent pipeline supporting emerging technologies. A new generation of CDN intelligent cache engines and algorithms will be implemented through a learning architecture. Enhancing cache algorithm business perception and data-driven AI self-learning capabilities, improving CDN hit rate and  enhancing CDN outflow, can greatly save CDN bandwidth and server costs. Research directions includes learned memory caching algorithms, learned disk caching algorithms, multi-level cache intelligent scheduling algorithms, intelligent cache admission algorithms, and cache portrait algorithms.

The research mainly focuses on the CDN + P2P architecture and algorithm of end-to-end cloud collaboration. The intelligent P2P content distribution network effectively disperses hot content, greatly reduces the peak pressure and bandwidth cost of CDN and improves the user experience. Research directions include P2P low-latency networking algorithms, P2P fast distribution algorithms, P2P hotspot prediction algorithms, P2P scarce seed prediction and scheduling algorithms.

With the rapid development of cloud video services, video has taken on new forms, with more and more abundant scenarios, higher resolution, stronger interactivity, and flexible access devices. Video encoding and decoding, as a core video technology, also faces challenges such as delay, bit rate, and power consumption. Supporting AI content awareness, combination of software and hardware, and flexible deployment to provide ultimate experience,  AI encoding is a research on AI+video codec technologies to provide video encoding and decoding algorithms  to various video service scenarios on the cloud，which has advantages of low bit rate, low power consumption, low latency and high quality.

As the largest traffic carrier of cloud vendors in the future, video is a competitive high ground with excellent experience and low costs. AI capabilities are required to continuously optimize algorithms and maintain the leading position of the algorithm. Research on cross-terminal lightweight AI enhancement algorithms, such as video reconstruction, super-resolution, super-frame, ROI, and quality enhancement, is beneficial to improve cloud video experience and competitiveness.

In storage and database systems, the cost of data storage needs to be reduced. Lossless compression algorithms can reduce data storage space without loss of information. Currently we are focusing on lossless compression algorithms for various types of data, including multimedia
data (images, videos), general data, time series data, etc.

Basing on machine learning, deep learning and other algorithm technologies, the algorithm provides multiple operational analysis capabilities around HUAWEI CLOUD's multiple business operation scenarios, such as cloud service product recommendation, business marketing management and control, and user churn prediction, etc., which would improve HUAWEI CLOUD's operational efficiency and user experience.

With the help of machine learning and AI algorithms, we are able to finish real-time analysis of risk relationship association and mobile device behavior, helping HUAWEI CLOUD operations resisting risk users strongly associated with violating users, and at the same time defending against group control device attack groups, targeting common Internet black and gray products. Fraud attacks form intelligent and efficient prevention and control, and build an operational security defense system.