The world's top experts and scholars in the field of artificial intelligence and industry people, in the human-computer interaction, machine learning, pattern recognition, industrial real-time artificial intelligence frontier topics for in-depth exchanges and discussions.
Dr. Deng Li, the chief scientist of Microsoft Artificial Intelligence, will be invited to attend the conference and will give a keynote report titled “Three types of deep learning models for driving multiple applications of big data artificial intelligenceâ€. On the eve of the conference, Dr. Deng Li received a brief interview with CSDN reporters to analyze big data, deep learning and other technical areas of artificial intelligence.
Dr. Deng Li introduced that his theme report will describe the relationship and difference between in-depth supervision learning, deep unsupervised learning and deep reinforcement learning, and the relationship between these three types of deep learning models and big data, and explain them through practical cases. The applicable environment and effects of the three types of learning algorithms. He believes that one of the limitations of current big data-based artificial intelligence is that it relies on deep supervision learning that requires large input and output of matching training data. The idea of ​​cracking the current lack of big data artificial intelligence includes deep unsupervised learning and depth. Reinforcement learning, as well as new structural characterization based on high dimensional tensors.
He also explained the application areas of reinforcement learning, the advantages of deep learning and reinforcement learning, and pointed out the role of deep reinforcement learning on Microsoft's Bot vision - deep reinforcement learning not only controls the content of each single Bot dialogue output, but also Control the coordination and switching between various Bots.
Deng Li, chief scientist of Microsoft Artificial Intelligence
He is a world-renowned expert in artificial intelligence, machine learning and speech language signal processing. He is currently the chief scientist of Microsoft Artificial Intelligence and the research manager of the Deep Learning Technology Center. He received his master's and doctoral degrees from the University of Wisconsin, USA, and then taught at the University of Waterloo in Canada. In the meantime, he also held research positions at the Massachusetts Institute of Technology. He joined Microsoft Research in 1999 and started the Deep Learning Technology Center in early 2014, where he led Microsoft's and Institute's technical innovations in artificial intelligence and deep learning. Dr. Deng Li's research interests include automatic speech and speaker recognition, oral recognition and understanding, speech-to-speech translation, machine translation, language mode, natural language processing, statistical methods and machine learning, neuroscience, auditory and other biological information processing, Deep structure learning, brain-like machine intelligence, image language multi-modal deep learning, commercial big data depth analysis and prediction. He has made significant contributions in these areas, including ASA (American Acoustical Society) Fellows, IEEE (American Institute of Electrical and Electronics Engineers) Fellows and Directors, ISCA (International Voice Communications Association) Fellows, and with deep learning and Outstanding contributions to the direction of automatic speech recognition won the 2015 IEEE Signal Processing Technology Achievement Award and the 2013 Best Paper Award. At the same time, he has published more than 300 academic papers related to the above-mentioned fields in top magazines and conferences. He has published 5 books, invented and co-invented more than 70 patents. Dr. Deng Li also served as editor-in-chief of IEEE Signal Processing Magazine and IEEE/ACM TransacTIons on Audio (Speech & Language Processing).
The following is an interview record.
It’s not surprising that AlphaGo beat Li Shishi
What are the current advances in technology research and application in the field of artificial intelligence? AlphaGo?
Deng Li: Based on my experience and understanding of the huge learning capacity of deep neural networks since 2009, I have no intention to make any progress in the current deep learning and technical research in the field of artificial intelligence and the large-scale successful promotion and achievements of various parties. It is not surprising that AlphaGo, led by deep reinforcement learning, beat Li Shishi four to one.
You appreciate the combination of intensive learning and deep learning, so in which areas is it suitable for intensive learning?
Deng Li: Reinforcement learning applies to any staged process of control and decision making, like chess, mechanical robotic action, and many business decisions. The best reward signal for applying reinforcement learning is to be clear or easy to define, such as playing chess. Otherwise, intrinsic motivation is used to drive reinforcement learning, such as for chat bots. One of the directions is to integrate information theory and dynamic programming, which is still in the research stage. The large state space used to be a big problem for reinforcement learning, but now the problem is basically solved after the introduction of deep learning. Intensive learning with large mobile spaces (such as the use of synthetic natural language as the "action" output of dialogue robots) is being studied in depth by our team. Using deep learning to solve large mobile space is much more troublesome than solving large state space. Our team has published some papers in this regard.
How do you see the combination of deep learning with more other methods (such as Bayesian methods) and prospects?
Deng Li: At present, deep learning based on neural networks can be well combined with Bayesian methods and generative models. The advantage is that it can give deep learning to explainability, and can also reduce the requirement of depth learning to match the amount of training data matched by input and output. If deep reinforcement learning is used, this combination can greatly improve learning efficiency because it makes the exploration steps in reinforcement learning faster and the exploration space becomes more extensive.
In addition, what technical areas do you think are completely unrelated to deep learning deserve our attention?
Deng Li: The reasoning of propositional logic and first-order logic seems to have nothing to do with deep learning. Recently, a lot of good work has been done with deep learning and logical reasoning. Our team has a long article in ICLR in 2016. However, using pure symbolic propositional logic and first-order logic to make reasoning is much easier to explain than using deep neural networks, just as high-level languages ​​of computers are much easier to understand than assembly languages. This interpretability is important in practical applications. But the logical methods and models of pure symbols are much more difficult to learn than deep neural networks. Fortunately, there is a set of evolving theories in cognitive science that establishes a purely symbolic tree or graph structure (which can be used for logical reasoning with high efficiency and strong explanatory) with a high-dimensional tensor. Structure. Because tensor is the most natural data structure for deep learning, this isomorphism allows our artificial intelligence system to effectively implement structure-to-structure symbol mapping (such as natural language or computer program input and output), but at the same time It is possible to directly learn and optimize this structural mapping using deep neural networks (this includes complex multi-step logical reasoning).
Microsoft artificial intelligence research and development route
Microsoft plans to become a leader in the field of artificial intelligence. Can you introduce the role of your work? What is the main progress of your work in the last six months?
Deng Li: I am currently managing and leading the Deep Learning Technology Center at the Institute of Microsoft Headquarters in Microsoft, 50% of the time. There is a strong technical and research team here. Another 50% of the time is in the commercial department of Microsoft as the chief artificial intelligence scientist, applying artificial intelligence and various deep learning technologies and research results to artificial intelligence products and cloud services. The main tasks of our team in the last six months include:
Successfully applied deep learning techniques to the analysis, forecasting, customer scoring of commercial big data, etc., and achieved remarkable results;
Promote the development of multi-class natural language dialogue robots with deep reinforcement learning;
Multi-media research and application combining natural language, vision and knowledge base;
Advances in basic research on new deep learning architectures, algorithms, and structural characterization.
Microsoft has done a lot of artificial intelligence APIs for developers. Do you think artificial intelligence will become a universal attribute of future apps? What artificial intelligence related knowledge does the current developer need to learn? How to get started?
Deng Li: Microsoft's CogniTIve Services (Microsoft Cognitive Services, its development documentation and tutorials can refer to the official website: https://TIve-services/) will provide more and more artificial intelligence tools to developers. Many have been called from the Microsoft Bot Framework. The Microsoft Bot Framework website https://dev.botframework.com/ has quite detailed information.
Regarding the ideals of Microsoft Bots, what major technologies do you think it needs to achieve?
Deng Li: Artificial intelligence with powerful functions is one of the most important aspects of Bots' ideal vision. The main technical accumulation is the deep learning mentioned above, especially the deep reinforcement learning. Deep reinforcement learning not only controls the best content of the dialogue output of each single Bot, but also controls the best coordination and switching between various Bots.
CCAI sharing big data and deep learning
Please briefly introduce your report topic "Three types of deep learning modes for driving big data artificial intelligence multiple applications" in this conference?
Deng Li: I mainly want to talk about the three types of deep learning models - the relationship and difference between in-depth supervision learning, deep unsupervised learning and deep reinforcement learning. Which model is used in which artificial intelligence application? why? Where are the insights? I want to use the success (and failure) examples my team has used to provide some insights.
I also want to talk about the relationship between these three types of deep learning models and big data. The large training data with matching input and output will generally make the depth supervision learning successful (end-end backpropagaTIon is effective for training with matching big data), but the input and output matching cost is very high. Conversely, the cost of big data without input and output matching is much lower. To train a deep learning system with large training data that has several orders of magnitude greater matching of input and output than the existing input and output, it is necessary to develop a new deep unsupervised learning algorithm. If successful, this will bring a new milestone to artificial intelligence.
Can you explain the application limitations of big data-based artificial intelligence and some ideas for cracking the high-quality big data artificial intelligence?
Deng Li: One of the limitations of current big data-based artificial intelligence is that it relies on deep supervisory learning, which means that end-to-end backpropagation can only be used after large training data with input and output matching. Not only is the cost high, but the system is very inflexible and it is difficult to adapt quickly to the new environment. If you need to solve complex logical reasoning problems, artificial intelligence systems based on big data and deep learning often do not give the ideal answer. This system that relies on deep supervisory learning lacks common sense and intuition.
The ideas for cracking high-quality big data missing include deep unsupervised learning and deep reinforcement learning as described above, as well as new high-dimensional tensor-based structural representation and knowledge base.
Who is suitable for listening to this report? What kind of preparatory knowledge do you need? What will they gain?
Deng Li: Researchers, graduate students, ICT companies and government managers. People interested in artificial intelligence and deep learning. I hope that after listening to this report, I will have a deeper understanding of artificial intelligence and deep learning, especially how to apply theory to practice.
Last question, how do you view the difference between foreign and domestic artificial intelligence technologies and applications? In terms of enterprise application and personnel training, what good experiences abroad are worth learning from?
Deng Li: The difference between foreign and domestic artificial intelligence technologies and applications is shrinking. Many of my former friends who have been engaged in artificial intelligence and deep learning in the United States and the United Kingdom have returned to China to start businesses.
Good foreign experience is to focus on innovation and encourage rapid failure.
50Hz To 60Hz Frequency Transformer,Audio Output Transformer,High End Audio Output Transformers,Toroidal Transformer Audio
IHUA INDUSTRIES CO.,LTD. , https://www.ihua-inductor.com