Founded in April 2010, Xiaomi is an Internet company that focuses on smartphones, smart devices and the Internet of Things platform. Xiaomi is dedicated to making amazing products that will touch your heart, and have honest pricing. Committed to continuous innovation on smartphones, smart devices, Internet TV and other products, Xiaomi makes positive adjustments to the enterprise layout, in AI, new retail, internationalization, fintech and so on. With the innovative advantages of its 3-pillar business model: hardware, new retail and Internet, Xiaomi rises rapidly and becomes one of the representatives of innovative “Internet+” companies.
In 2017, only in business for 7 years, Xiaomi grossed over a 100 billion in revenue and broke records in the global business industry. Xiaomi was also listed on the main board of the Hong Kong stock exchange on July 9, 2018 (1810. HK). In 2019, Xiaomi became the youngest company on Fortune Global 500 List and are the fourth Chinese Internet Company being shortlisted after Jingdong, Alibaba and Tencent.
On Jan 11th 2019, Xiaomi’s co-founder and CEO Lei Jun announced at the annual review meeting that it officially launched the “mobile + AIoT” dual-engine strategy this year. In the next five years, Xiaomi will continue to invest more than 10 billion RMB in AIoT (AI + IoT). On August 2019, Xiaomi was selected as Smart Home National AI Open Innovation Platform. In the future, Xiaomi’s AI will base on this platform and focus on 5G+ AIoT, and make more products and applications more intelligent.
Facing the extremely strong business needs of hardware enabling, mobile Internet, IoT, e-commerce, finance and others, Xiaomi established its AI, big data, and cloud platform departments quite early. AI technologies have been widely used almost everywhere in Xiaomi’s smart devices, e-commerce sites and internet applications. The AI department is one of the core departments, and includes AI Lab, XiaoAi Tongxue, AI Ecology, AI Virtual Assistant and other teams. The AI department builds AI products and provides business insights through analytics. Currently, the AI department has more than 600 researchers and engineers, focusing on six directions including Computer Vision, Acoustics, Speech, NLP, Knowledge Graph, and Machine Learning.
The main research interests of computer vision team include image processing, image understanding, human face detection, video processing and understanding, etc. The main goal of the team is to help Xiaomi phones create the ultimate photo shooting experience. And thus, the team did research on the basic image quality algorithm of the camera, the intelligent editing, recognition and understanding of image and video. In recent years, Xiaomi phones has contributed the core algorithms to the key functions: AI camera, single camera bokeh, face beautification, face unlock, face album, magic changing sky, smart frame picker, photo album text search and so on. In addition, the team also applied different algorithms to TV, smart devices and other business scenarios and met the business needs. In terms of fundamental research, the team ranked 1st in FDDB face detection competition and published a Deep Exposure paper in NeurIPS 2018.
The acoustics team is committed to technical research and engineering in the field of AI acoustics and speech enhancement, providing intelligent acoustic basic algorithm to Xiaomi phones and IoT full ecological hardware products, and building a leading intelligent voice interaction experience in industry. At present, 6-microphone array, nearby wake-up, cooperative playback, sound processing and other technologies have landed in a number of Xiaomi products, and at the same time, the first fully automatic far-field acoustics lab has been built in China. Xiaomi far-filed acoustic test specification has been widely accepted by the industry as soon as it is launched, and it was confirmed as the acoustic test standard of China AI industry Alliance, which has established Xiaomi’s leading position in the field of far-field acoustics.
The main research and development directions of the speech team are speech recognition, speech synthesis, speech wake-up and voiceprint recognition. The main goal of the team is to develop or find key voice technologies for the company in key voice scenarios. At present, all of these technologies have been landed on Xiaomi products such as Xiaomi phones, TVs and speakers. As the products landed, the team also published several academic papers on top tier conferences, such as Interspeech and ICASSP.
Natural Language Processing (NLP) team mainly focuses on fundamental NLP techniques, Spoken Dialogue System, Machines Translation, Text Generation and so on. Fundamental NLP technologies include Segmentation, Part-of-Speech Tagging, Semantic Parsing, Emotion Detection, Text Classification, and Intent Recognition. Keeping up with the cutting edge of technology, NLP team built Xiaomi MiNLP platform, optimized models based on business needs, and provides with NLP services supporting local, cloud and mobile deployment. Dialogue System supports XiaoAi Tongxue MiChat function, using BERT pre-train model, deep matching and text generation techniques, achieves much better performance compared to baseline models. Machine Translation optimized distillation model, creatively used quantitative computation, and thus achieved higher model compression ratio. XiaoAi Laoshi supports multi-lingual offline translation, and its Chinese-English translation is very accurate.
Knowledge Graph (KG) team investigates on the KG construction and applications. Now the KG includes 13 industries (categories), in total 3 billion data records (triple), and had been widely applied to Question- Answering (QA), Customer Service, Advertisement, Information Feeds and other products. While developing best practices, KG team combined data-driven and knowledge-driven approaches, and achieved pretty good feedbacks in different product lines. KG team also summarized the methodology of how to effectively construct KG. After optimized the Named Entity Recognition, Entity Disambiguation algorithms, the quality of generated KG triples had been enhanced. The KG team also developed the general knowledge and specific domain QA system: general knowledge QA system answers millions of answers from XiaoAi Tongxue users everyday, and the specific domain QA system helps with customer service and sales assistant scenarios.
Machine Learning team is committed to build up a world-class Machine Learning and Computing platform. The Mobile AI Compute Engine (MACE) framework focuses on deep learning inference on devices. According to the heterogeneous computing feature of multi-vendor chips, MACE did performance optimization, from the aspects of user experience, ROM usage, power consumption, model encryption, etc. Now MACE is open-sourced, and has gained great attention. MACE won the 13th China, Japan and Korea Open Source Software Technology Excellence Award and Big Data Expo Outstanding Project Award in 2019. The AutoML group focuses on Neural Architecture Search (NAS), and proposed MoreMNAS, FALSR, FairNAS, MoGA and Scarlet models, which applied to different business scenarios, and achieved better performance than state-of-the-art models. And ClouldML platform provides deep learning environment, GPU computation capabilities, GPU training tasks scheduling services, to maintain and develop relevant systems.
XiaoAi Tongxue is an AI product supported by AI department and other internal teams. It is a voice interaction enabled platform, and had been built inside the smartphones, speakers, TVs, watches, and other 20+ devices. In the meantime, XiaoAi Tongxue is able to “communicate” with more than 190 million Xiaomi IoT smart devices, and directly control more than 39 categories, more than 590 types of smart devices. It can not only turn on the lights, adjust the air conditioning temperature, control the sweeping robot, but also report the weather, play the music and tell stories. In fact, XiaoAi Tongxue is almost good at everything. Now the MAU of XiaoAi Tongxue exceeds 49 million, and total number of wakes reaches 20 billion. XiaoAi Tongxue is one of the most active voice interaction enabled platforms, as well as the largest smart device control platform in China.