From AI School to AI Lab, Microsoft has been making efforts to promote the popularization of artificial intelligence and lower the threshold of AI development. According to Dr. Zhou Lidong, vice president of Microsoft Research Asia and head of the Joint Center for Artificial Intelligence Systems, the key to breaking the current bottleneck of deep learning lies in the system. The highest level of the system is "completely invisible." To this end, Microsoft Asia Research Institute is working hard to create an open source project for deep learning-deep learning intelligent exploration Neural Network Intelligence (NNI), through the easy-to-use "invisible" system, intelligent and automatic learning to help more development By.
At the 2017 Microsoft Build Developer Conference, Microsoft Senior Project Manager Cornelia Carapcea demonstrated a technology:
The custom vision API model created by the user only needs one training data sample (only dozens of photographic samples), while Custom Vision can complete the rest.
Once the model is created (just a few minutes), the user can access it through the REST API installed on the Microsoft server. Carapcea says it can be used to identify food and landmarks, even in retail environments.
Custom Vision can select the image that is most likely to add the greatest gain to the model, allowing the user to manually mark the image, and then continue to improve the overall accuracy and reliability.
In an interview with AI frontline reporters, Dr. Zhou Lidong, vice president of Microsoft Research Asia and head of the Artificial Intelligence System Joint Center, said: Microsoft Custom Vision and Google Cloud AutoML are both an application of AutoML (automatic learning) in the field of visual recognition. Both are very simple and easy to use. By hiding the complex model selection and parameter tuning process behind the product, Microsoft Custom Vision can provide users with a very simple and easy-to-use experience, enabling non-professionals to customize their own models with zero threshold.
At the same time, Microsoft Custom Vision is also a part of Microsoft Cognitive Services. The cognitive services launched by Microsoft include more than 20 APIs in five categories: vision, speech, language, knowledge, and search, such as face recognition and emotion. Recognition, speech recognition, spell checking, language understanding, etc. The focus of Microsoft Cognitive Services is to provide users with a universal service. Users can directly call ready-made smart APIs to develop smarter and more attractive products without spending a lot of time training models by themselves.
Regarding the recently emerging AutoML technology, Dr. Zhou also expressed some of his own views:
In recent years, "AutoML (automatic learning)" has become a research hotspot. In an automated way, the machine tries to learn the optimal learning strategy, thereby avoiding inefficient manual adjustments by machine learning practitioners. Classical AutoML methods include Bayesian optimization (Bayesian OpTImizaTIon) for hyperparameter adjustment, and meta-learning technology (Meta learning/Learning-to-Learn) for optimizer and network structure adjustment. In addition to arousing widespread research interest in academia, AutoML has also been practically applied in the industry. For example, the Custom Vision service provided by Microsoft Azure mentioned earlier can facilitate cloud computing users to automatically train for computers. Visual machine learning model.
AutoML allows users to use machine learning without professional knowledge to greatly reduce the threshold, or even zero threshold. Without the guidance of machine learning experts, users can get high-quality models under certain circumstances through AutoML, which makes the industry application of machine learning easier and feasible. Dr. Zhou said that Microsoft Research Asia hopes to make the use and research of such technologies more popular. Microsoft is working hard to create a project for deep learning, called: Deep Learning Intelligent Exploration (Neural Network Intelligence), referred to as NNI.
The purpose of developing the NNI toolkit is to provide users with an intelligent and automated deep learning toolkit that can help users or developers automatically analyze data, automatically help them search for models, perform parameter debugging and performance analysis, and automatically iterate through this During the preparation process, users can save more time and focus on exploring more in-depth machine learning.
Dr. Zhou Lidong said that NNI integrates the AutoML algorithm, which is a toolkit that supports different operating systems and can be run locally in the cloud. It is especially suitable for scientific researchers with a certain artificial intelligence foundation to select more targeted and accurate models. Microsoft has developed a brand-new language for this, making it only need a few lines of code to define and describe the search space, and all the complex issues such as the underlying communication are encapsulated, which is completely transparent to users. NNI focuses on solving a series of system problems that support AutoML, and effectively accelerates the innovation of AutoML algorithm researchers in this field in an open manner. In order to encourage more artificial intelligence researchers to strengthen this research together, NNI plans to open source in the near future.
Another way to build an intelligent system
However, Microsoft's idea of ​​lowering the threshold of artificial intelligence and increasing the popularization of deep learning does not stop there. AutoML technology is only one part of it. What Microsoft wants to do is an intelligent system.
Dr. Zhou said that AutoML is part of the artificial intelligence system, and it also poses a series of challenges to the artificial intelligence system including resource management and task allocation. Simply put: a good, flexible and extensible artificial intelligence system can better support AutoML and make it better and faster to generate results. A good AutoML can make the artificial intelligence system more complete and more convenient to use.
Intelligent era, system first
Dr. Zhou Lidong first said: If artificial intelligence has no system, it is just a mirage-it looks beautiful, but it is not true at all.
He believes that the system is to make complex things orderly and easy to use. In the computer field, the importance of systems is self-evident. In the entire computer development process, every major breakthrough we feel is actually driven by many computer system theories and designs.
Dr. Zhou analyzed the importance of the system in each era:
In the Internet age, there are many very new systems appearing in our lives, and one of the most representative systems is the search engine. Many of you use search engines to search for information on the Internet, and there are many systems and theories behind search engines. One of the typical system theory is the distributed system theory.
In the era of big data or cloud computing that everyone is familiar with, cloud computing systems are very typical systems. Now we see that many companies are providing cloud computing services. Here are some new computer system technologies, including virtual machine technology, fault tolerance technology, etc., because these technologies make such services and systems possible.
In today's artificial intelligence era, we see more and more deep learning breakthroughs in the fields of computer vision, speech recognition, natural language, etc., which creates greater demand for systems. We have researched and developed many large-scale deep learning platforms, which also rely on recent system progress and research results, including how to use heterogeneous hardware to efficiently perform these deep learning tasks, and how to perform high-performance Parallel Computing. These make deep learning, especially the processing of very deep models possible.
Dr. Zhou said: "We can also imagine that quantum computing will become a mainstream technology in the future. But we can also see that the development of the system has not yet reached the stage where it can be put into practical use. Although the theory is very mature. , But if it is to become a reality, a lot of innovation, research, and practice in the system are needed to bring about this big change."
The bottleneck of deep learning
In his speech, Dr. Zhou mentioned that deep learning has achieved further development, but it will still encounter some bottlenecks, and many bottlenecks are still generated in the system. He said: "Now, even if we have many different hardware accelerations and many different models, how can we map this model to the corresponding hardware very efficiently and have different customized optimizations? The entire deep learning field? The work inside is actually done manually, rather than done in a systematic way."
In addition, when training deep learning models, developers generally use GPUs for acceleration. When the training samples are only one million, a single-card GPU can usually meet our needs, but when the number of training samples reaches tens of millions After hundreds of millions of levels, single card training will take a long time, so at this time, it is usually necessary to use multiple machines and multiple cards for acceleration. In this case or some more complex cases, we need a lot of system design and consideration. Practitioners of artificial intelligence no longer need to worry about whether they need to build a GPU cluster themselves to do artificial intelligence things. Because this is what the system should accomplish, Dr. Zhou said: "We should do all of these things, then artificial intelligence researchers can be freed to focus on artificial intelligence problems."
The threshold of artificial intelligence is very high, so the value of artificial intelligence practitioners is also very high. This is also because a lot of work at the system level is not deep enough. Dr. Zhou said that by advancing their work, they hope to lower this threshold and truly achieve the "popularization of artificial intelligence." This will be a very feasible and necessary next step.
Microsoft's Intelligent System
Dr. Zhou believes that the most critical innovation of the system is to achieve perfect abstraction. Users actually "cannot see" the system. When everyone talks about it, it seems that they see breakthroughs in the upper level such as vision, and never see the progress of the lower level. Therefore, we have always said that the highest state of the system is completely invisible and invisible.
He told us: Microsoft has always believed that in the future, the whole world is a computer. Whether it is the real world or the virtual world, all these parts will be connected together.
Dr. Zhou said: "The understanding of Microsoft Research Asia in the field of systems is: First of all, the system is a cross-domain and wide-ranging research. We need to apply from the underlying hardware to the upper level, including new applications such as artificial intelligence. , Have a very broad understanding, so that we can design a system suitable for applications. Secondly, we also need to have a deep understanding of the principles of compilers, compilation optimization, database systems, so that we can make full use of them through reasonable system design The ability of the hardware, that is to say, the research on the system must also be deep, which is also very important."
According to reports, the research results of Microsoft Research Asia in the field of systems have been widely used in real business scenarios and have produced actual business effects, such as:
GraM distributed parallel graph processing engine, capable of processing graphs with more than 1 billion edges in memory through clusters;
The Apollo big data task scheduling system can be directly deployed on 100,000 machines and can schedule millions of tasks every day to support the daily business of search engines, advertising and other departments;
The StreamScope distributed streaming data processing platform can process billions of advertising information in real time;
The KV-Direct Key/Value system can handle more than 1.2 billion operations per second, which is at least an order of magnitude improvement over the most cutting-edge research systems of its kind. Although this system has not yet been deployed, it is a representative of the forefront of Microsoft's current research, published in the top conference on computer systems SOSP 2017. These are just the tip of the iceberg. It is said that Microsoft's exploration of intelligent systems goes far beyond this. Microsoft in the intelligent age seems to have more possibilities.
The future of intelligent systems
Finally, Dr. Zhou expressed his personal views on the development trend of intelligent systems:
First of all, many deep learning frameworks will be interoperable and unified in the future. In fact, the database is a good example-a long time ago there were various databases, but in the end everyone invented the so-called RelaTIonal algebra (relational algebra) based database, so that all database models become a unified model . In terms of artificial intelligence, this unity of intercommunication is also something that will definitely happen from a system perspective.
Secondly, the capabilities of the system will become stronger and stronger, and there will be unbounded resources. The whole world is a computer, so the final goal we hope to achieve is that in an environment with unbounded resources, the resources you use, whether from a computing center or from your own computer, or even from someone who doesn’t know Whatever it is, it will be well hidden by the system. You just need to do things well and don't care about where the resources come from.
Finally, there is a point that needs to be emphasized again. The most critical innovation in system research is to propose a more concise abstraction, and to support this abstraction with new tools and platforms, so that everyone's work efficiency can be improved.
end
In order to promote the popularization of artificial intelligence and lower the development threshold, Microsoft is constantly working hard and trying, and I believe that users will soon be able to enjoy the results of these efforts. As Dr. Zhou Lidong said: The highest state of the system is invisible. Perhaps the highest state of technology is intangible. This is how good technology changes our lives imperceptibly. Although we don't feel it, it does happen.
The 2018 Microsoft Education Summit with the theme of "System|Accelerating Future Change" will be held from August 1 to August 3, Beijing time. As a top-level dialogue platform recognized by industry and academia, this conference will bring together academics from all over the world Experts from the world, Microsoft and Microsoft Research have jointly discussed how to build today's "AI supercomputer" composed of "intelligent cloud + intelligent edge" covering the world. Click on the two posts to learn about the highlights of this summit and the online live broadcast schedule.
Optical Rotary Sensor,Custom Encoder,Optical Encoder 6Mm Shaft,Handwheel Pulse Generator
Jilin Lander Intelligent Technology Co., Ltd , https://www.jllandertech.com