Intel Labs China (ILC), one of three "regional research institutes" set up by Intel Intel Institute abroad, currently focuses on researching key technologies related to future autonomous systems, including artificial intelligence algorithms , Independent system platform and intelligent infrastructure three parts.
Song Jiqiang, president of Intel China Research Institute, said: "Intel China Research Institute is committed to creating a world-class industrial research institute that will create Intel's innovation engine and maintain Intel's technological leadership with groundbreaking results. It not only looks forward to the future revolutionary Science and technology research, but also carries the mission of transforming the research results into global and Chinese indigenous innovative products.We must act as pragmatic explorers as well as bold practitioners.For the future, we will always uphold Intel's promotion of Moore's Law over the past 50 years To speed up the promotion of innovation, promote the development of industry and society, and allow the progress of science and technology to benefit every citizen of the world. '
Artificial Intelligence Algorithms Related to Visual Comprehension: Making Your Machine More Understand Your research interests include face recognition, emotion recognition, scene understanding, and deep neural network compression based on advanced AI algorithms developed by Intel China Research Institute Visual understanding and visual content creation capabilities, to create an unprecedented experience.Front-looking visual understanding, visual perception mainly involves three aspects: face analysis and emotion recognition; for visual recognition efficient visual compression; visual, language and Knowledge Fusion In this open day event, Intel China Research Institute experts focused on the latest technological advances in three areas.
1. Face analysis and emotion recognition technology
For face analysis and emotion recognition, Intel China Academy had a very early stage of algorithms and won the EmotiW Race (Emotion Recognition Contest in Natural Context), which was the second team's 200 Times, and can be applied to driverless scenes. Based on Intel Real-Time Portrait Style Migration System, Intel China Academy real-time migration of a variety of artistic portraits to static target images and dynamic target video. The technology can be widely Applied to AR / VR, online live, film and television, multimedia social networking and other fields.
In addition, Intel China Research Institute has developed Intel 3D face facial expression capture technology based on machine learning and depth learning algorithms.Relevant researchers from Intel hundreds of thousands of face image data, training multiple face recognition models for video images Intelligent analysis and 3D modeling, not only can detect and identify faces, but also accurate reconstruction of 3D faces, and real-time tracking changes in facial expressions, pre-designed effects of the material attached to the 3D face and superimposed on the video, In order to achieve realistic cool facial effects.
Deep neural network compression technology
At present, one of the major challenges facing deep learning is that mainstream DNNs are computationally and storage intensive, which has led to huge challenges in the deployment of edge and embedded devices.Therefore, Intel China Research put forward from the dynamic Network Surgical DNS, Progressive Network Quantize INQ to MLQ Quantize Low-Precision, Deep-Squared Compression Solution Gains 100x DNN Model Lossless Compression Performance Based on the AlexNet test results, this concise solution Beyond at least double the mainstream depth compression scheme and over 100x network compression at 2/4-bit resolution.
3 multi-modal image video resolution
In recent years, how to automatically generate video description has aroused widespread interest of researchers.It is hoped that when a computer sees a video, it can "tell a story" according to the content of the video.However, the training of video description model usually requires a large amount of complexity and With some subjective artificial annotation.During the current dataset construction process, the annotation staff will read a video, the content of the video in one sentence.However, a video usually occurs several different events, Because of the subjectivity of annotation staff, people do not know which event it describes, nor which region on the different frames it describes.
To this end, Intel China Research Center first proposed the weak supervision of video-intensive description generation method, do not need to train the data to video different events and corresponding regions were marked separately, and only use the label of a video description, you can Automatically generate multi-angle video description, and pick out the most representative description, so that the computer does not have to 'teach people', but can do 'give top priority'.
Autonomous system: to create a robot's 'heroic' era
Autonomous system is a research direction related to artificial intelligence.The goal of Intel China Research Institute is to launch a set of autonomous intelligent robot development platform based on CPU + FPGA heterogeneous computing to help researchers develop based on this platform.The reason why the introduction of heterogeneous Platform because robots research needs to introduce both time and space computing models.CPU as a perfect representation of time computing model, instructions executed in chronological order, can achieve flexible control; accelerator is more like a spatial computing model, The use of spatial parallelism to complete a large amount of data processing at one time, so that complex calculations are accelerated.
With this goal in mind, the Heterogeneous Extensible Robot Open Platform came into being as a set of low-powered, low-power tools designed by Intel China Research for intelligent robots including service robots, medical robots, autonomous vehicles, Consuming, high-performance, small-footprint heterogeneous system platform solution.In this scenario, the CPU as a control center, with FPGA and other dedicated accelerator chip (such as Movidius's VPU) to provide highly efficient performance.The complete set of HERO hardware system The Intel Core ™ Series CPU, powered by the Intel Arria 10 GX Series 1150 FPGA, is a heterogeneous accelerator that can process large amounts of data in real time and run a variety of intelligent algorithms.
At this open day event, Intel China Research Institute announced a new partnership program around the HERO platform around which Intel China Research is committed to working with the industry's platform, application and ecosystem partners through R & D Cooperation, academic innovation and industry demonstration, together to create a complete, efficient and open platform, continue to expand the application scenarios, accelerate the arrival of technology and products.
Intelligent Infrastructure: Accelerating the Convergence of Communication and Computing This study focuses on supporting technologies that will ensure the smooth operation of large-scale intelligent networked devices in the future, including wireless communication technologies and intelligent storage technologies. The former studies MIMO technology and mobile edge computing at the network access end Technology that studies technologies that use Intel's latest 3D storage technologies and intelligent solutions built with FPGAs to provide high-throughput, low-latency, value-added service potential.
At this Open Day event, Intel introduced how to create a high-quality VR experience with the convergence of communications and computing for the next generation of 5G communications networks, and today's high-definition experience needs to be implemented to meet the needs of high-quality virtual reality experiences 25-40 times the resolution increase.To this end, Intel hopes through the integration of computing and network, end to end network expansion to create a remote and immersive interactive experience.
Intel demonstrated ultra-high-resolution panoramic video capture system, which is mainly used for wireless VR devices, users can watch panoramic video immersion, or participate in VR game interaction.It uses the ISP bus synchronization camera shutter, through the distributed architecture processing Sample images, and at the same time, use the optical flow method to spliced the panoramic video and interpolate the stereo information.The system also uses Intel Core GPU high-performance compressed images to wirelessly transmit the host rendering output to the remote terminal, and use the OFFLOAD technology to reduce the display delay. These technologies allow television viewers to experience the spectators' experience in the front seats of the Olympic Games even at home, or choose different viewing angles within the venue.