Shulou(Shulou.com)06/02 Report--

Http://blog.sina.com.cn/s/blog_cfa68e330102zg2l.html

2018-11-16 12:02:32

The GIAC Global Internet Architecture Conference will be held in Shanghai from November 23 to 24. GIAC is a technical architecture conference launched by the high-availability architecture technology community for architects, technical leaders, and high-end technology practitioners. This year's GIAC has been attended by Intel, Tencent, Alibaba, Baidu, Ant Financial Services Group, Huawei, iFLYTEK, Sina Weibo, JD.com, Qiniu, Meituan Dianping, ele.me, Caiyun, Geling Deep Mirror, Databricks and other company experts. Enjoy a 12% discount on tickets purchased this week, and as low as 40% discount for members of the highly available architecture.

On the eve of the conference, Gao Yafeng interviewed Deng Yafeng, the producer of the 2018 GIAC big data & AI Sub-Forum, and conducted an interview on big data & AI, which is of widespread concern.

Deng Yafeng, the current Chief Technology Officer of Geling Deep pupil Information Technology Co., Ltd., graduated from Tsinghua University, has 16 years of research and development experience in computer vision and artificial intelligence. In his past work, he has published more than 10 papers and applied for more than 100 Chinese patents, of which 95 have been authorized. He has worked in Baidu Deep Learning Research Institute, in charge of face recognition direction, and has led the team many times to achieve excellent results in mainstream face detection and face recognition competitions. Its main interest is to focus on how artificial intelligence, especially computer vision technology, has landed on a large scale in the real world from a technical, product and commercial point of view.

High availability Architecture: nice to interview you. You are a veteran in the field of AI. Can you briefly introduce you and your original intention of choosing to pursue a deep career in the field of AI? Why take a fancy to the development direction of AI?

Deng Yafeng: in 2002, I graduated to Tsinghua University as a graduate student. At that time, many students chose communication, which was a hot direction at that time. I chose AI by accident, but later I thought it was because I really liked this direction. I felt that it was very interesting and cool for the computer to understand the content in the image and video or to recognize what was said in the voice, so I chose this direction.

Started to do this direction, mainly out of interest and love of technology, feel that doing AI is a very challenging and interesting thing, so I came in. Before 2012, AI had always been an unpopular direction in industry because it had very few landings, and there were very few companies doing it at that time. I can persevere, on the one hand, interest and love play a great role, on the other hand, I believe in the value of AI technology and its great impact on our future lives from the very beginning. AI can be seen as an extension of industrial automation, helping people to liberate from simple and repetitive work, enhance human ability and improve human efficiency, so that we can live a better life and have a better world.

Highly available architecture: Gering's deep pupil is well known in areas such as computer vision. What do you think is the biggest difficulty in computer vision? What is the biggest difficulty encountered in the process of landing?

Deng Yafeng: although computer vision has made great progress, it is still difficult to truly produce, and it often needs to be adjusted and adapted to the scene. at present, the main challenge is how to develop standardized products that can be replicated on a large scale in terms of performance indicators, cost and support scale. In the process of landing, like other artificial intelligence technologies, the biggest difficulty is that there is no 100% match between the level of technology that can be achieved and everyone's needs. For example, in the past many years, the demand for face recognition has been there, but the technology can not be fully met, so it can not be landed. Even today, the technology is still not perfect, but we can not wait for the technology to be fully mature before landing. Therefore, the main challenge at the current stage is, based on the existing imperfect technology, how to integrate the factors of market, product and technology, turn technology into products and services to produce value as soon as possible, and use the revenue and data generated by landing to help talents, technology, products and markets form a virtuous circle.

High availability architecture: data plays an important role in the field of AI. Model training usually requires a lot of data. What is the general method of data tagging in the industry? A human flesh mark? How does GE Ling do deep pupil?

Deng Yafeng: due to the characteristics of deep learning models, current model training often requires a lot of data. For example, face recognition often requires billions of data. To label so much data in a violent way, on the one hand, the cost is very high (it usually costs a few cents to label a picture), on the other hand, many data labeling tasks are more difficult than people can do. We generally use the semi-automatic method to process the data, use the existing algorithm model to preprocess the data, and then mark the part that the machine can not do well but the human can do well by manual labeling. Through the internal relationship between the data to greatly reduce the labeling workload and improve the standard quality.

Highly available architecture: mathematics is the foundation of artificial intelligence. What mathematical knowledge do you need to learn to enter this field? Do different artificial intelligence directions need to master different algorithms? What algorithms are generally included in face detection and face recognition?

Deng Yafeng: mathematical knowledge is very important for students engaged in artificial intelligence. The current artificial intelligence is mainly based on statistics. At the same time, there are many operations related to derivation and matrix in the neural network, so, it will be helpful if you can master the knowledge of higher mathematics, probability and statistics, random processes, linear algebra or matrix theory. Of course, if you have mastered all kinds of knowledge of machine learning, the above mathematical knowledge is not necessary. And artificial intelligence is a subject that needs hands-on experience to master, so it is not necessary to learn the above courses well before you can start the study of artificial intelligence.

Different directions of artificial intelligence, because there are different tasks, there are different domain knowledge, so there need to be different models or methods to express, the algorithm will naturally be different. For example, the detection task in vision is less used in other fields, while many sequence-related models are used in speech recognition and semantic understanding, and less in the field of vision. But now, the methods of speech recognition, natural language understanding and machine vision are becoming more and more unified, and methods in different fields will learn from each other. Therefore, as an algorithm engineer, it is best to master some of them.

In the early stage, there are relatively independent algorithms for face detection, such as boosting, but in the era of deep learning, face detection algorithms and object detection algorithms are basically the same. At present, many general object detection frameworks are used, such as SSD, R-FCN and so on. The whole process of face recognition is divided into several modules: detection, feature point location and feature representation, of which the most important feature representation module needs to use the identity information of training data to learn universal face representation. in addition to network structure design, the most important thing is to design loss function. Early feature representation has softmax, contrastive-loss, triplet-loss and other methods in the loss function. Many of the improved algorithms of softmax are slowly becoming mainstream.

High availability Architecture: for people who want to master or improve their AI skills, can you give them a summary of the quick learning route?

Deng Yafeng: for students who have no foundation, the learning steps I suggest are:

Step1, read the basic knowledge of deep learning, and understand the basic concepts such as neural network, SGD optimization method, loss function, etc.

Step2, take a moment to familiarize yourself with a training framework, tensorflow, pytorch, mxnet

Step3, find a task to do, even if it is a simple classified task of MNIST. If you encounter a problem, check the information or ask someone for advice. In short, learning by doing is the most efficient. After completing this task, find a more difficult task to complete.

For students who want to improve, on the one hand, it is to find a task that is more challenging to their own abilities, and on the other hand, a very important shortcut is to join a very strong team. It is far better to learn and improve in a team than to fumble.

High availability architecture: what professional abilities should high-level artificial intelligence talents possess? What do you think the average artificial intelligence engineer needs to improve in order to become an expert?

Deng Yafeng: in industry, high-level artificial intelligence talents need to be very strong in algorithm ability, engineering ability and understanding of industry and products. They should not only see the big trend and the value of technology, but also know how to polish technology into products through algorithms and engineering. If ordinary artificial intelligence engineers want to be promoted to experts, they first need to improve their algorithmic and engineering capabilities, expand their horizons and technical fields, and gradually improve their understanding ability in the industry and products.

High availability architecture: the significance of deep learning to computer vision is self-evident, and now the former has basically become the standard of the latter. however, on the one hand, deep learning is very dependent on large-scale data; on the other hand, in many real-world application scenarios, it is often difficult to obtain large-scale data, in your opinion, how to find a balance between deep learning and the scale of data?

Deng Yafeng: the dependence of deep learning on data is determined by the way of optimization in the process of model learning. if you want to change fundamentally, you need to make a great breakthrough in the way of optimization, which is difficult in the short term. In practice, there are some ways to reduce data dependence, such as using transfer learning ideas, using data from other fields to train the base of models, and then using a small amount of domain data to learn, such as using semi-supervised or unsupervised methods to make use of a large amount of unlabeled data. In addition, we can also consider making use of the constraints of the problem to strengthen the constraints on the model, improve the generalization ability and reduce the dependence on data. Of course, data augmentation is also a very important way to increase the amount of data. Under the current technical conditions, in industry, how to obtain a large amount of labeled data at low cost should still be the first method to be thought of.

High availability Architecture: as the co-chairman of GIAC and the producer and lecturer of AI, you will analyze your experience in building large-scale computing systems from the perspectives of algorithms, data, computing and so on. In your opinion, what are the difficulties in building large-scale computing systems? What's the point? And what is the practical significance of large-scale computing systems?

Deng Yafeng: the ultimate goal of machine vision is to make all kinds of vision sensors intelligent, understand the characteristics, identity, behavior and relationships of people, cars and things in the physical world, and digitize the physical world. So far, no real large-scale visual computing system has been produced. The difficulty and key to building a large-scale visual computing system is that, on the one hand, we need to constantly improve the accuracy and scene adaptability of the algorithm in order to produce acceptable errors in massive data; on the other hand, we need to constantly improve the efficiency of the algorithm and reduce the cost, so that the processing time and cost for massive data can be bearable, and the third is to cooperate with big data technology to mine the relationship between cross-sensor targets. Once we can really build a large-scale visual computing system with acceptable cost, it will have a great impact on security, management, and business operations in many areas, such as smart cities, smart commerce, and so on. so that the offline world will be able to operate in a more intelligent and digital way, and the offline world and the online world will merge and open up, generating more business models and values.

It involves a lot of technology-related content, which I will expand in detail on GIAC on the 23rd. I hope my explanation can help you understand the cutting-edge technology, key factors and application direction of machine vision. I hope you can take fewer detours.

High availability Architecture: as co-chairman and important guest of GIAC, what kind of message do you have to the conference?

Deng Yafeng: I hope that GIAC will become a stage for technical personnel to communicate deeply and collide ideas, and bring real harvest and value to the participants.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.