Following "Let the machine understand your voice," Titanium invited six Titans to discuss how to make the machine understand the world. According to Zhang Hongxin, an associate professor of the State Key Laboratory of CAD&CG of Zhejiang University and a consultant of Rokid scientists, the book was organized in Titanium.
Professor Zhang is a Ph.D. student at Zhejiang University and a postdoctoral fellow in the Department of Computer Science at Hong Kong University of Science and Technology. He has conducted visits and cooperation research at Microsoft Research Asia, Aachen University, Germany. In recent years, more than 30 papers have been completed, many of which have been included in SCI/EI/ISTP and have achieved a high rate of citation. At the State Key Laboratory of CAD&CG of Zhejiang University, the theory of digital geometry processing, computer graphics hardware acceleration, 3D reconstruction, visualization and computer vision was systematically studied. The discrete differential equations and convolution theory were successfully applied to fast three-dimensional modeling. Cooperated with Alibaba Cloud to develop a rendering cloud system and participated in the development of the Rokid family companion robot.
The following is Professor Zhang Hongxin's sharing in Titanium:
Hello everyone, I am Zhang Hongxin from the National Key Laboratory of CAD&CG of Zhejiang University. I am also a consultant to the Rokid Robot R&D team. Thank you to the audience. I also thank the Titanium Media for giving me such an opportunity to work with many of the experts in the artificial intelligence industry. This sharing.
Although I have been doing machine learning related application research for more than ten years, I am still not a real expert in machine learning or artificial intelligence. I can barely be a veteran player because my own research interests are mainly in graphics. The field of intersection of learning and computer vision was later extended to visual analysis of data. Taking this opportunity, I want to reflect the thoughts of our generation through some of my personal experiences, and then tell me some rough understanding of artificial intelligence for everyone to discuss.
Our generation belongs to 70. I remember that one of my favorite cartoons when I was the youngest was "Astro Boy". The earliest sci-fi movie should be the "Star Wars" trilogy. At that time, R2D2 in the movie. The robots are particularly fascinating, and these films have made our generation have the earliest intuitive understanding of smart robots. But to be honest, I didn't know what "artificial intelligence" was at first. I just thought it was science. These cute robots will one day appear in our lives and become good friends of our humanity.
Coincidentally, until one day two years ago, my good friend, the founder of Rokid, found me and said, "We do robots." I didn't hesitate to say "Okay, let's do it together!" So, Start this fun exploration journey with Rokid's team. In the process of this exploration, I found that there are many research problems that are promising, so now I am happy.
seedIn retrospect, from primary school to middle school, with the increase of age, I really like to read a few magazines, namely, "Science Pictorial", "UFO Discovery" and "Mystery". I also like the physics and mathematics behind it. Then, if you are a little bigger, you like to watch Radio. However, it was later discovered that many simple small productions in "Radio" are OK, but complicated production requires a lot of instruments, which cannot be afforded for a junior high school student or a high school student. So I feel that it is better to take a step back, because mathematics is not bad, just use mathematics as a basic tool to explore the world. Later, I was also very fortunate. I was admitted to Zhejiang University, and I was a student of the Department of Reading. That is, in college, I had a chance to meet Missa, because he liked playing guitar like me. I didn't expect two people to work together for many years because of playing guitar.
In the third year of college, probably around 1996, there was a very interesting thing. At that time, the head of the mathematics department was Professor Chen Shuping. He once asked us to chat with a group of students and asked us what we were interested in. When it was my turn to speak, there were three words in the mind - the robot. Then I said to Teacher Chen, I really want to be a robot. Teacher Chen was stunned for a while, smiling and encouraging me to say, “I need to learn more professional knowledge in the future, and cooperate with the teachers who are automated and computer.†Perhaps it is because of the encouragement of the teacher, which later prompted me to spend a lot of time. Learn computer technology and minorize in electrical engineering in partial engineering.
Therefore, these experiences during the university have planted the seeds of today's 70-generation generation of these smart devices and smart technologies, and now have Rokid's products. It is for this reason that we are willing to jump out of our own small circle and combine elements of electronics, computer technology, and many mathematical algorithmic logic. To be honest, I like to play guitar and love music. When we were able to gather together, we could explore this somewhat deviant but somewhat cross-border beauty.
It is precisely because of this, I think artificial intelligence, can be simply solved as a set of algorithms to form the automation logic, it is a combination of software and hardware technology, only physical material is king.
Small data eraLater, I was very fortunate to successfully send a master's degree from Zhejiang University and directly to a doctoral degree. The title of the paper he gave me at the time was "Study on Discrete Methods of Modeling and Rendering Complex Shapes." In this topic, we actually study a spline theory, trying to express a three-dimensional scene through a procedural subdivision surface method. So at the time, I extensively explored how various 3D models can be expressed and how to effectively draw them.
During my Ph.D., around 2001, I went to Microsoft for an internship for a while. Joined the vision group led by Dr. Shen Xiangyang of Microsoft and worked as an intern on his side. At that time, I came to this group and gave me a special question. I hope that I will do research on relighTIng technology. What is heavy lighting? Specifically, it is to study from one picture or multiple pictures, firstly to restore the three-dimensional scene, and then further guess the material properties of the calculated object according to the information of the three-dimensional scene (complement: finally, re-lighting the three-dimensional scene). At the time, this was a difficult problem, and it required a lot of computer vision techniques, including image segmentation, stereo vision, and many methods of data statistics. These methods can be regarded as a form of artificial intelligence, and also some typical methods in machine learning. At the time, at Microsoft, there was a very good atmosphere. I remember that there were a lot of very powerful people at the time. Now they have become researchers and professors, such as Sun Jian, Liu Ce, etc., and everyone has discussed a lot of methods of machine learning.
My work at Microsoft belongs to the field of "visual computing." The direction of visual computing is actually the intersection of computer graphics and computer vision. These two directions can be said to be a natural pair. Computer graphics is a process of generating two-dimensional images from three-dimensional scenes or data. Computer vision is the reverse, which is the result of guessing or predicting three-dimensional images from two-dimensional images. Especially for the content of image-based reconstruction, it can be said that graphics and computer vision are inseparable. Therefore, in the academic world, the two are added together as "visual computing."
What impressed me the most during Microsoft was that I was exposed to the so-called small sample learning method from that time. Because we had a discussion class at the time, such as Sun Jian, Wang Tianshu, Liu Ce, and other friends, including Zhu Xichun, everyone discussed various visual methods together. During this period, we focused on the study, a very famous book written by Dr. Vapnik, the inventor of SVM, called "The Essence of Statistical Learning Theory." In this book he has extensively studied the theory of the SVM method. The purpose of the full text is to learn to obtain a good statistical model by comparing small samples. This model is obtained by calculating the data. Under such an atmosphere at the time, in fact, everyone is exploring a variety of machine learning methods, in addition to the SVM method, there are some more important techniques for dimensionality reduction.
After graduating from the Ph.D., I worked as a postdoctoral fellow at the Hong Kong University of Science and Technology. The teacher who worked there was Professor Dai Qiulan. She worked with her on sketch-based human-computer interaction. At the time, Hong Kong University of Science and Technology gathered a large number of researchers from all corners of the country. Some of them were post-doctoral there, and there were also teachers there. During this period, I was honored to meet Teacher Zhang Zhihua. Everyone kindly called him "Lao Zhang". He is now teaching at the Jiaotong University and Peking University. Lao Zhang is both my teacher and my friend. We often walk together at the seaside of Hong Kong University of Science and Technology. When he was walking, he told me a lot of statistical methods. To this end, we have done some data reduction work in the follow-up cooperation. Finally, one of the algorithms is used for model decomposition and texture mapping in graphics. At that time, everyone especially praised the statistical methods of small samples, and used a lot of theoretical and computational techniques in statistics. But at the time everyone was more scornful of the neural network method, which is a very interesting process.
From about 2005 to around 2006, I had the privilege of visiting the Aachen University of Germany, where the professor was Leif Kobelt. At that time, he mainly did a lot of research on 3D mesh processing. Many of his team's technologies were actually exported to BMW cars, so there are many horizontal issues. During my visit to Professor Kobelt, I focused on the further introduction of machine learning techniques into graphics for the processing and analysis of 3D models. Because at the time, I had a very interesting job with a doctoral student at Zhejiang University's CAD lab called Xu Dong. Our technique can interpolate between different 3D models. The computational theory behind it is the differential method on the mesh, which we call "Poisson shape interpolation".
The method of Poisson shape interpolation, Professor Kobelt is also very interested, because he found that this technology may be used in the shape design of BMW cars. Later, we gradually realized that this technology may be an essential expression of an object. But how to verify this is hard to say because it requires a lot of data. Therefore, starting from this project, I have been paying more and more attention to some progress in the graphics database. In this case, we found a lot of data, which is actually suitable for some data-driven methods. If you can accumulate more data, you can not only do shape analysis and search of graphics, but also drive some related physical simulations.
I remember one year, I worked with a doctor named Song Chao to do a technique about physical simulation. In this simulation technology, we not only use some shape data collected in advance, but also incorporate a physical mechanics model. We combine the two to achieve a more realistic, but very fast physical simulation.
But after doing this series of research at the time, to be honest, we are actually somewhat confused in terms of academics, especially in the direction of graphics. The direction has grown to a very high altitude and entered a platform period. So we are making some attempts and want to do some mining from other angles to see if there is any other way to go. This is precisely the time when we introduced people in the 1970s to the era of big data.
Big Data Era
Probably around 2010, we did three different projects before and after, and I summarized these three projects into three “bigâ€.
Big calculationAround 2010, Dr. Wang Jian of Alibaba Cloud found us and hoped to render our existing graphics calculations. This very complicated calculation process was moved to Alibaba Cloud. Later, after two to three years of hard work, the project became the first Saas application on the line.
On one occasion, our two sides cooperated to mobilize a total of 6,500 computers, or computing nodes, to complete the rendering task of an animated film called "Kunta Legend", which was shot by a local advertising company in Hangzhou. It is very remarkable that we dispatch such large-scale computing resources to do such a complicated computing task. It is because of the ability to mobilize such computing resources on a large scale, so that I can say that it is open, and I am very excited, I feel that this kind of big computing power can really do something.
Thin Film Module,Acrylic Wall Panels,Plexiglass Panels,Perspex Panels
zhejiang goldcity technology co,ltd , https://www.membrane-gc.com