The "big model" of tackling, the way to look for the technical system "unified"

At the 2022 Viewing Technology Open Day (Megtech 2022) event that ended a few days ago, the co -founder and CEO Yinqi said that the AIOT core technology research system of "2" is to support the success of the future to the success of AIOT commercialization. Important cornerstone.

Among them, the AI ​​technology system composed of "basic algorithm scientific research" and "scale algorithm production" is an important part of this set of AIOT core technology scientific research systems.

For a long time, the work of "basic algorithms" has greatly promoted the "emancipating mind, realizing cognitive upgrades and technical breakthroughs". It is not difficult to find out the development of computer vision. The breakthroughs of each generation of basic models have greatly promoted the development of visual AI and promoted the application of algorithms in more scenarios.

At the perspective, the burden of "basic algorithm scientific research", deeply digging the "research, development and deployment" of "basic models" fell on the shoulders of Zhang Xiangyu, head of the basic model group of the Institute of Vibration and his team members.

In response to the work of re -tracing the root causes to achieve cognitive upgrades, Zhang Xiangyu admits, "A good basic model improves the performance of the entire system. How can we design high -speed, high -precision, low -power basic models? It belongs to its own scientific taste and research methods, and it is important to continue to achieve cognitive breakthroughs and upgrades. "

Emancipating the mind to open the road of innovation

So far, Zhang Xiangyu's transcript has shined enough. He published more than 50 papers in top -level conference/journals such as CVPR/ICCV/ECCV/NIPS/TPAMI. The number of Google Scholar references exceeded 170,000, and continued to export and efficiently output on neural network models such as ResNet, Shufflenet, and REPVGG.

Under his leadership, the basic model group went up a year in a year, and the small goal of "one person per person per person" in the group was also achieved smoothly. CVPR 2022, they put forward the CNN and MLP design paradigm based on large Kernel, including the dynamic convolutional neural network FOCAL Sparse CNN; in addition, they will also publish the preliminary work on the new network PETR perception of autonomous driving perception of new network PETR work. On ECCV 2022. Zhang Xiangyu emphasized that "PETR has almost no artificial design component, but is completely based on the same architecture processing multi -view, time, multi -tasking, and multi -modal input."

At this time, these technological innovation work that had a pivotal position in the field of neural network research, could not help but marvel at the prediction ability of his research team's "bleak stroke" and the forward -looking vision of the development of the commercial world. But if everything is simply attributed to luck, it is obviously unsuccessful. In fact, thanks to the guidance of Dr. Sun Jian, he and his friends around him always tried to find those "anti -intuitive" pioneering cognitions, solidified it into knowledge, and eventually precipitated into technical beliefs.

"Once you find that one thing that I never thought about is what can be done, this can often bring pioneering results." Zhang Xiangyu took the academic community for the controversy of Transformer and CNN. The difference between transformer and CNN is better. But we see the same thing behind the two. We think it has little to do with whether it is transformer or CNN. It is important that its feelings are not great. This also shows that compared with the ability, the optimization characteristics of the model architecture are often more important. "

Continue to think down along this idea. It is not difficult to find, "Once the model is unified, the design of the AI ​​accelerator will be very simple, that is, one model can be suitable for various equipment and various tasks, but the challenge it brings is also significant For example, to share a model and a algorithm on multiple tasks, we must deepen this system and understand this model in order to abstract the commonality, and then use a unified model to achieve the past specifically for all for all for all. The performance of the system alone has the performance. "

The advanced perception of this is the prelude to the way to start the road of scientific and technological innovation.

Reading literature development research system

Where does "anti -intuition" come from? In fact, the major discovery of Shi Po Tian is not uncommon in the scientific research community. Many "new things" are just new bottles of old wine. It is another indication of some phenomena that has been discovered in the past. The thesis, found that the predecessors have done research "is always the most headache for scientific researchers.

From the perspective of the team members of the basic research group, it is a minor probability event to touch the air. The fundamental method is to change their knowledge priorities and constantly change their thinking.

Zhang Xiangyu highly respects the literature and archeological research methods of Professor Ma Yi, the Department of Electronic Engineering and Computer Science of the University of California Berkeley. "Teacher Ma Yi will always go to the chain of the literature to find a certain idea from which literature. Although many documents have discovered some facts, a paper usually only transmits limited conclusions, and the author may not have consciousness. The meaning of this fact is in other circumstances. Find the 'point' that the existing knowledge system cannot explain, dig in depth the relationship behind these facts, try to explain it in their own language and organize them organically To form your own technical belief and research system. "The" pearl "scattered in different documents requires a lot of effort. At a round table forum that was discussed around the RACV 2021, Zhang Xiangyu posted a very rigorous brain map to guide the audience to follow his thinking.

There are more than one scholar in each box to discuss the papers, but he followed the demonstration process of the demonstration process, quotation, and arguments of all the articles along the demonstration process outside the conclusion, and then analyzed some of the torrents of the words. Converse facts.

Rao is so, "with your own thoughts, do something different", you still have to go through various tests at the practical level. Artificial intelligence technology continues to evolve forward, and the correction and reflection of its own technical beliefs has become daily skills. This must be comprehensively collecting information and clearly what the current technical limit is done. Find the answer to the answer. " As Zhang Xiangyu said, "Some key technologies have always been a step by step, and there are always risks to choose a technical route."

At the just -concluded 2022 Viewing Technology Open Day, Zhang Xiangyu clearly pointed out that "big" and "unity" are a new trend of basic research on visual AI. In this regard, he emphasized that the "big" definition definition is to give full play to the power of big data and high calculation with the power of innovative algorithms to expand the boundary of AI cognition; Expressing various data and tasks with modeling will get simple, powerful, and universal systems.

Adhere to the original and the ideal work

As the "disciple" of Dr. Sun Jian, Zhang Xiangyu admitted that his scientific research taste, scientific research values, scientific research mentality, and even team communication and cooperation ability were almost all learned from "Boss Sun". Because of this, he always believes that "can think independently and refuse to follow the trend; strong prediction ability, dare to shoot research lines; solid basic skills, clear how to do the symptoms", etc. Literacy.

The scientific research atmosphere of the basic model group may just be regarded as the actual ground version of these abilities. Following the main task logic of computer vision, the research direction of the basic model group focuses on the four aspects of universal images, large -driving models, computing large models, and video understanding. Select one in the mode.

The project system has a clear time node, which will regularly review and track progress. It also needs the group to make the group forces to solve the problems that arise; the free exploration is based on the team members' own interests to give full play to the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to carry out the subjective initiative of the team members to conduct Title choice. Zhang Xiangyu assumes the responsibility of "setting direction" and "details" in the group, but he said that his more important responsibility is to maintain the atmosphere of his favorite things, stimulate the creation of the big guy force.

After entering the vast ocean of deep learning, it has been in a hurry for 10 years. According to the path that the seniors pioneered, the former young students have eventually grew into a scientific researcher who can form their own technical beliefs. Not long ago, Zhang Xiangyu decided to officially rename the Base Model group to the Foundation Model group. The difference in the only word reflects the ambitions of the basic model group to commit to the development of large visual models.

At the end of the speech of the open day of the open technology, Zhang Xiangyu said that basic scientific research will always adhere to the original, practical and essential scientific research values. "Only by achieving originality can we break through the cognitive boundary of existing technology. Only by achieving practicality can we truly transform scientific research results into products and the value that can be landed. Only by discovering the essence, we can from the complicated appearances. See the innovation point behind the model, better realize the "big" and "unified" basic model research and development. "

This statement is in line with the concept of "technical beliefs, pragmatic values" pursuing the "technical beliefs, pragmatic value" of the Institute of Vibration. Ten years of grinding swords, "the fire of the stars can be alarm."

Finally, for young people who are interested in engaging in computer vision research, Zhang Xiangyu also gives four dry goods suggestions based on his own experience and in -group conditions:

Broadly accumulated knowledge. Massive reading literature is extremely important. "As far as the world -renowned scholars I have come into contact with, none of them are not amazing reading. Many people now do not read scientific research. All members must participate in the "Paper Reading" once a week and submit an interpretation report on time.

A keen problem consciousness. On the basis of reading a lot of original literature, we must have the ability to summarize knowledge and find problems. "A valuable papers must be explained according to my knowledge system. I will record it. When you look at other literature in the future, once you encounter similar or opposite cases, you will reflect on it. In the end, because the experiment is wrong or implied the details I didn't realize before, or here is a new cognition. "A solid mathematical foundation. The solid mathematical foundation can increase the upper limit of AI research, but it is difficult to make a long time after graduation after graduation. Therefore, he encourages students to work hard to lay a good basis for mathematics. In order to prevent forgetting the knowledge points, Zhang Xiangyu will re -brush the textbooks at the undergraduate stage every six months to maintain the feeling.

Pure research mentality. Because the anxiety caused by the pressure of the output of the papers is the most important reason for the vast majority of people to give up scientific research. But the interesting thing about scientific research is that they never understand to understand, and never know to know that the paper is only a by -product of this process. To maintain the original intention of scientific research.

- END -