Doubt and glory, technology and business, all the questions, perhaps through such a domestic independent AI framework, at least half can be answered.
Hit the charts, stunts, controversy, money burn
Innovation, genius, star, critical thinking
…
When evaluating AI companies, some people say that they are money-wasting machines and shared bikes of the technology industry.
Others say that they are the light of Science and technology in China, gathering the most academic talents and killing the most industry records, bringing something that only stayed in science fiction movies at first into everyone’s life in just a few years.
But no matter how we evaluate it, we have to admit that without AI, China’s tech industry would be half the star it is today.
Without the presence of megvii, Sensetime, Yitu and Yuncong, China’s AI industry will lose half of its star.
In any case, it may not be the valuation, nor the ranking, that is responsible for the continued success of an industry that is accelerating with all kinds of top talent.
Technological innovation and product landing are the two most important dimensions of test value, and the landing and innovation of a real AI enterprise must be supported by the existence of an efficient and flexible ARTIFICIAL intelligence algorithm platform.
Take megvii as an example. At the business level, based on its Brain++ artificial intelligence algorithm platform, megvii has successively developed a series of products and services directly applied to commercialization, such as AIoT digital software platform, dongjian smart city management operating system, hetu robot network operating system, etc.
In terms of technology, based on Brain++, megsight has topped the list of major image recognition competitions for many years, and innovatively proposed the latest scientific research achievements including MegDet, ShuffleNet, AutoML (Single Path one-shot NAS), etc.
Even, megvii official once summed up the significance of its Brain++ : it is “Brain++ that makes large-scale algorithm training possible.”
So take Brain++ as an example, how did megvii build such an AI algorithm platform that supports both research and development and landing?
Support Brain++, and how some black technology figure?
Meanwhile, looking to the future, what evolution will Brain++ produce?
Most importantly, in today’s open source algorithm everywhere, through Brain++, we can see the existence of AI enterprises and what is the meaning of?
With controversy and glamour, technology and commercialization going hand in hand, it is never easy to read them. The AI self-research framework represented by Brain++ is an excellent sample to observe in its progress.
After five years, megvii’s Brain++ ambition and growth
While looking up at the starry sky of technology, while stepping on the ground of commercialization. This is perhaps the most romantic gesture for a true AI company.
Behind the creation of this romantic gesture, it needs a constantly evolving and blossoming algorithm platform to carry out scientific research and landing support.
Speaking of Brain++ ai algorithm platform, Dr. Jian sun, chief scientist of megvii and president of megvii research institute, described its importance to enterprises as follows: it is “Brain++ that makes large-scale algorithm training possible. **** “. Meanwhile, with Brain++, megair “is able to tailor a rich and growing portfolio of algorithms to the fragmented needs of different verticals, including many long tails. In addition, we can develop new algorithms with fewer people and in less time.”
It is no accident that Dr. Sun jian gives such a high evaluation of Brain++. After all, megvii started to build such an engine for its AI research and development — Brain++ around 2014, shortly after its establishment.
Looking around the AI industry at the time, everything seemed to be in its infancy.
That year, training an AI model took at least one or two months, and developers even had to type C++ by hand to complete the calculation process.
That year, there was only one industry-wide deep learning framework, Theano, launched in 2007 with near-heritage status, and Caffe, launched in 2013.
Caffe’s emergence as a first-generation deep learning framework, however, is significant to both the industry and the academic world. However, Caffe’s design, as an early-generation deep learning architecture, still has some limitations.
For example, Caffe’s network structure is defined in the form of configuration files and lacks relatively free, flexible and visual algorithm expression represented by computational graphs. ** As of 2015, some large neural networks, represented by ResNet with 152 layers, have emerged, and Caffe becomes very cumbersome to use for such large neural networks.
Therefore, megvii Brain++ established the idea of building the framework in the form of computational graphs at the very beginning of Caffe’s architecture.
Looking at the industry, this graph-based approach is one of the key features of the second generation of AI frameworks, such as TensorFlow and PyTorch. However, the difference is that Brain++ appeared half a year earlier than TensorFlow, which also has more targeted advantages and affinity for visual image developers in megvii.
In addition, future interviews provide a glimpse of Kuang’s ambition and commitment to building a fully autonomous and usable AI architecture.
In a 2017 Series C funding interview, Megvii founder Inch said the industry “needs a better engine like TensorFlow,” when talking about his AI platform. An All in One system, that’s what we’re working on the most.”
Up to now, sun jian, president of megvii research institute, told us that Brain++ has been gradually upgraded to version 8.0, which has great advantages in both stability and technological advancement.
At the same time, megvii’s thousands of r&d staff are all using its internal Brain++ for technology development. Megmegdet, AutoML and other technologies have also grown from the tree based on Brain++.
Thus, a huge algorithm empire is clearly showing us its endless power source and infinite possibilities in the future.
Deep interpretation, black technology and ecological construction behind Brain++
Dismantle the composition of Brain++, which contains a total of artificial intelligence data platform, artificial intelligence foundation (training) framework, and artificial intelligence computing platform three major parts of the content, ** corresponds to the three major elements of AI development: ** data, algorithm and computing power.
According to Sun jian, Megvii can greatly increase the speed and cost of data annotation by allowing algorithms to work together with data and using algorithms to assist in data cleansing and annotation.
The specific method is to first train a primary algorithm with a small number of annotations, and then use this algorithm to annotate data. Then, for the uncertain data encountered in the algorithm annotation process, the most uncertain data is manually annotated, and the algorithm automatically annotates the remaining data again. Thus, more efficient and faster annotation results can be obtained with the minimum time and cost.
In addition, in terms of data, Megvii and Beijing Zhiyuan Artificial Intelligence Research Institute jointly released Objects365, a data set containing 700,000 pictures, which contains common objects in 365 and more than 10 million labeling boxes, making it the largest object detection data set in the world at present. This is ten times larger than the previous largest COCO dataset.
The ai infrastructure framework in Brain++ can simply be thought of as a completely self-developed TensorFlow or PyTorch.
As a typical second-generation AI development framework, Brain++ is more flexible and definable than its predecessor, Caffe. In addition, Megvii combines its autonomous megvii AutoML technology to automatically optimize the neural network for different platforms, further improving the flexibility of the algorithm.
Users can log in to the computing platform as if they were logging in to a virtual machine. And there is no need to submit tasks after debugging, as long as the required CPU/GPU number is automatically applied, and you can stop to debug at any time in the process of training, find out the error and then continue training.
In addition, multitasking and multi-user scheduling capabilities are also a feature of Brain++. According to the principle of transparent computing, Megvii has developed a layer of software for computing resource management and scheduling optimization, which can temporarily recover resource allocation when users are idle, thus supporting hundreds of researchers to perform hundreds to thousands of training tasks on tens of thousands of Ggpu at the same time.
The combination of the two methods can maximize the use of computational power while ensuring the training efficiency.
How is Brain++ different from Caffe or TensorFlow?
The first is completely independent research and development. Sun jian said that the significance of Brain++ is to better adapt to their own research and development needs, otherwise for TensorFlow and other frameworks to “change the existing things, in fact, it is difficult to change or inefficient. It’s hard to make these changes quickly when the code is already large. With our own stuff, you can easily modify and validate things.”
Second is customized optimization for computer vision tasks: in Brain++, megvii has done a lot of optimization for computer vision.
Take the CVPR article published by Sun Jian in 2018 as an example. This article provides a new detection called “MegDet” for accelerating deep neural network training from the perspective of mini-batch. This technology for the first time realizes that a “large Mini-batch” detector with more than 256 samples can be used to train tasks with 128 GPU cards during training object detection, thus reducing the training time from 33 hours to 4 hours.
But in addition to these black science and technology, in fact, set up such a large collection, work force, the data, the algorithm framework, which integrates AI platform, it is difficult not only from technology and capital investment, greater difficulty is how to from the start of business founded set up such an open, but not long-term ecological forecast returns.
An industry expert gave us this analogy:
The difficulty and significance of an enterprise to do a good job in AI platform can be compared to huawei’s intention to build a Hongmeng operating system + ecosystem.
And to disassemble its difficulty, the first step in front of it is how to build it, which is a very huge project.
Second, there is the issue of time. Starting early not only means more time to build from the bottom step by step, but also means that you do not have to face a pile of accumulated years of huge code base, just like the soft installation has been arranged, only to find that the reinforced concrete structure at home is not built well.
The most important is the ecological construction, how to gather high-quality developers, how to ensure the efficient use of the framework, even for internal use of enterprises, this is still an unavoidable problem.
Brain++, the glory and future of megvii
Based on Brain++, what exactly does megvii do?
Sun jian says all of megvii’s research so far is based on Brain++. The most intuitive performance is that, for two consecutive years in 2017 and 2018, megvii won the most authoritative international image recognition contest based on Brain++.
At the algorithm level, megvii has also developed deep neural networks that can be deployed in cloud, edge and mobile terminals, based on Brain++.
In the cloud, in 2015, the team led by Sun Jian when he was working in Microsoft first proposed the deep residual network ResNet, which made it possible to train hundreds or even thousands of layers of network, machine vision beyond human eyes, and greatly improved the performance and ceiling of cloud algorithms.
Edge side: In 2016, Megvii developed a lightweight deep neural network model DorefaNet suitable for chips. Through low-precision methods, convolution computation can be completed on chips with simple bit operations.
On the mobile computing platform, Megvii developed ShuffleNet, a lightweight convolutional neural network, in 2017. Through point-by-point group convolution and channel mixing, ShuffleNet allows the use of more feature mapping channels under the condition of determining the computational complexity budget. Therefore, ShuffleNet realizes the design of encoding more information on lightweight network, which can be specially applied to mobile devices with limited computing power.
Since 2017, ShuffleNet has developed ShuffleNet V1, ShuffleNet V2 and ShuffleNet V2+ versions: On ARM-based mobile devices, ShuffleNet achieves speeds up to twenty times faster than classic AlexNet with similar accuracy. And has a 741 star rating on GitHub.
In comparison to MobileNet, the web-optimized architecture developed by the Google AI team (according to the Google AI team’s public technical reports), ShuffleNet V2 is often 30-50% faster than MobileNet V2 in terms of actual performance.
GitHub: github.com/megvii-mode…
In addition, Sun jian also introduced megvii’s AutoML technology — “Single Path One-Shot NAS”.
For a long time, there will be a joke in the industry called the so-called artificial intelligence, there is as much intelligence as there is artificial. But AutoML allows machines to automate end-to-end optimization, dramatically reducing labor costs.
Previously, the research in the field of AutoML has been “monopolized” by foreign enterprises and platforms such as Google AutoML Vision, Microsoft Custom Vision and Amazon SageMaker. Domestic Baidu EasyDL is also conducting relevant exploration in the past two years. Megvii is the current leader in this field.
In terms of application, Megvii has launched a project called ultra Quality in the industry.
Through deep learning technology to replace the original ISP process, kuang can see testing, analysis of the original image, such as noise reduction and fusion processing, able to solve the user in the night and low light environment photograph when the screen brightness darker, excess noise, dynamic range, poor quality problems, such as, and through the study of artificial intelligence imaging characteristic of high quality digital camera, Restore the original details of the scene texture, so that the overall quality of the picture is improved.
But a big problem with the new generation of algorithms is that they are very computationally intensive, which requires a trade-off between computation and effectiveness.
According to Sun Jian, Megmegv can automatically optimize the neural network for different platforms through AutoML, so as to retain the most efficient operators, and reduce or discard some inefficient operators without affecting the overall effect, so that the algorithm can play the highest effect in the corresponding platform.
The picture on the right is the result of ultra – resolution processing
However, sun jian also said that he can give an 80 grade to Brain++ as there is still a lot to do looking ahead.
Megvii, for example, has been planning a proprietary language for deep learning training since last year to reconcile the flexibility required for training with the performance required for reasoning. At present, leading enterprises in the field of deep learning and deep framework are making some pioneering explorations, but in terms of progress, we are still in the early stage.
Conclusion: understand Brain++, is not only the AI framework, but also the counterbalance of Chinese industry under the node of The Times
Before talking to Sun Jian, I had rehearsed in my heart the possible death posture of countless AI enterprises: compete with EACH other in AI technology, and maybe Ali Cloud, Tencent Cloud and Baidu can crush them with their capital and scale advantages; But when it comes down to it, traditional industry players who have mastered open source technology seem to have the best market and experience.
But what is the point of an AI company? Perhaps in addition to down-to-earth things like landing technology, we also need to look at the stars.
Based on a solid AI algorithm platform, we can constantly explore some unknown algorithms and technologies with bumpy road ahead. Layout of some previously unknown languages and applications. At the same time, it is constantly evolving, all the way back to the most appropriate user needs of the application.
Perhaps, if the time is extended to 50 or 100 years, all the stars and enterprises we look up to, question and forget will eventually be completely different, but there will always be something left, perhaps an algorithm, a language or a flexible framework or an open ecology.
Furthermore, in the field of chip, operating system and even browser, we have been under the control of others for a long time, but in the emerging field of AI, we have managed to start with our rivals and surpass them at multiple nodes. Isn’t it the existence of such a silent but far-reaching AI platform and framework as operating system?
Brain++, represented by Brain++, as the autonomous and controllable algorithm framework of a few Chinese AI enterprises, not only affects the pace of their own scientific research and commercialization, but also directly determines the confidence of the whole country’s industry to compete with overseas in this new era.
What is the future of commercialization? Hit the list, stunt, controversy, burn money; Innovation, genius, star, critical thinking?
The answer lies in the minds of everyone who is in the midst of this change and everyone who is watching.