Shulou(Shulou.com)06/02 Report--

2019-11-25 10:09:09

As an enabling technology, artificial intelligence has been gradually applied in many industries, but at this stage, artificial intelligence is still in the stage of technological improvement in some aspects. This paper will focus on the academic research of artificial intelligence and briefly discuss the academic frontier and technological development trend of AI.

Chen Junlong, Vice Chairman of China Automation Society

Hot Technology Research topic of artificial Intelligence

At present, artificial intelligence has been widely used in many industries, but in terms of popular technologies, there are still some defects in security and other aspects. Using simple "adversarial patch" (adversarial patch) can make face recognition ineffective in reliability, security and so on. This also makes AI technology research need to pay more attention to artificial intelligence "attack" and "defense" topics, how to ensure the high security, reliability and robustness of AI algorithm, artificial intelligence algorithm in security "attack" and "defense" is also a very hot technology research direction.

Another research topic is to generate countermeasure network (GAN, Generative Adversarial Networks). There are many applications of generating countermeasure network, such as "image generation" and "data enhancement", such as face recognition image generation in the field of security. Generation countermeasures network is also gradually used in other application fields, such as material design in life science, as well as food and drug design, which has also become a new research direction of artificial intelligence technology.

Edge computing is also one of the focuses of artificial intelligence technology research. With the advent of the era of the Internet of things, more and more data are generated at the edge, and edge computing has become a significant trend. At present, there are many ways of edge computing to determine what should be put on the edge side. At present, the mainstream aspect of artificial intelligence algorithm is to put the algorithm on the cloud and the reasoning part of programming on the edge, which can ensure that the process of data operation and training, which requires high-power operation, is carried out in the cloud, while the reasoning process takes place at the edge. However, this mode may lead to the lack of intelligence of the algorithm at the edge end, so how to improve the intelligence of the algorithm at the edge side is an important research direction.

The birth of the wide learning network is considered to be a good substitute for the original deep learning network. It is designed based on the idea that the input is randomly mapped into an extended node as input (according to the early random vector function linked neural network, random vector functional-link neural network). It can calculate the correct mapping very quickly, and it can also learn in real time in the incremental part of the data. According to a large number of test results, the real-time application effect of width learning is very good in the process of edge computing.

The bottleneck of artificial Intelligence in Theory

Academician Tan Tieniu once said that there are still great limitations in artificial intelligence at this stage, which can be summed up in four sentences: intelligence without wisdom, IQ without EQ, accounting can not be "calculated", and there are professionals without generalists. Explain specifically the bottleneck problems that are limited in theory mainly include:

(1) data bottleneck: deep learning requires a large amount of data; (2) generalization bottleneck: this is a common problem faced by pattern recognition, computer vision and artificial intelligence methods. the existing methods still can not achieve the ideal generalization performance in some practical problems, or the generalization performance of the trained model decreases obviously in the changing environment or domain. (3) Energy consumption bottleneck: although the human brain is a general artificial intelligence system, its energy consumption is very low (only 20 watts), but the existing artificial intelligence systems implemented on computers consume high energy. (4) semantic gap bottleneck: at present, most language services are simple queries, do not involve semantic reasoning, and lack real language comprehension, such as some ambiguous natural language sentences. it is easy for people to understand their true meaning according to context or common sense, but it is difficult for computers to understand. (5) interpretable bottleneck: the existing artificial intelligence systems know what it is but don't know why, they rely too much on training data and lack of deep-seated data semantic mining. Therefore, interpretability is very important, artificial intelligence should not only know why, but also know why it is only shallow intelligence, know why it is called deep intelligence; (6) reliability bottleneck: the reliability of existing artificial intelligence systems is poor, some misrecognition results can bring fatal consequences, especially in the field of autopilot.

The Technical Research Direction of artificial Intelligence

In order to better break the bottleneck of the above artificial intelligence technology research, in the field of AI academic research, some new technology development trends and research directions are also worthy of attention, including:

(1) it is a general trend from dedicated artificial intelligence to general artificial intelligence. Some technology giants, including national institutions, are laying out the research of general artificial intelligence. Microsoft has set up an artificial intelligence laboratory to challenge general artificial intelligence as the main goal.

(2) the interpretable artificial intelligence system has attracted much attention, and it will become an important direction to break through the bottleneck of statistical learning. DARPA's report: the first wave is based on rules, such as a series of methods and techniques represented by expert systems; the second wave is the current statistical learning represented by big data drive; at the same time, they think that the third wave is likely to be interpretable artificial intelligence, that is, artificial intelligence needs to know why, so as to see the important significance of artificial intelligence interpretability.

(3) small sample or even zero sample learning has become an important direction to improve the generalization ability of artificial intelligence system. The recently proposed generation countermeasure network, capsule network and generation model are all beneficial attempts to reduce the demand for training data and improve the generalization ability of artificial intelligence systems.

(4) non-depth neural network computing model has become an important direction of machine learning innovation. At present, the theoretical foundation of deep learning is weak, the model structure is single, the resource consumption is too high, and the data dependence is strong. Non-neural network and resource-saving machine learning model is expected to become the next breakthrough.

(5) brain science and artificial intelligence are deeply integrated, coordinated and complement each other. IBM's TrueNorth chip, DARPA's MICRONs research project and some brain-inspired intelligent computing models recently published in Science and Nature journals are typical examples of this.

Research and Application of width Learning (BLS) Network

In addition to the above-mentioned key points of AI academic research, there are also wide learning (BLS) networks that deserve more attention. Width Learning (BLS) has been widely studied and applied in scientific research institutions (Chinese Academy of Sciences), domestic well-known universities and enterprises since it was first proposed by us (Professor Chen Junlong and his team) in 2018.

Although deep learning networks are very powerful, most networks are plagued by extremely time-consuming training processes. First of all, the structure of the deep network is complex and involves a large number of hyperparameters. In addition, this complexity makes it extremely difficult to analyze the deep structure in theory. On the other hand, in order to achieve higher accuracy in the application, the depth model has to continuously increase the number of network layers or adjust the number of parameters. In order to overcome these problems, the width learning system provides an alternative method for deep learning of the network. at the same time, if the network needs to be expanded, the model can be reconstructed efficiently through incremental learning.

In the aspect of design idea of width learning (BLS), firstly, the features of the input data mapping are used as the "feature nodes" of the network; secondly, the features of the mapping are enhanced to randomly generate weighted "enhanced nodes"; finally, all the mapped features and enhanced nodes are directly connected to the output, and the corresponding output coefficients can be obtained by the pseudo-inverse of express delivery (or gradient descent method). The most important feature of BLS lies in its single hidden layer structure, which has two important advantages, one is "horizontal expansion", and the other is "incremental learning". Different from deep neural network, BLS does not adopt the structure of deep neural network, but is based on single hidden layer neural network, and can use "easy to understand mathematical derivation to do incremental learning".

To put it bluntly, the deep neural network learning architecture begins to learn after the structure is fixed, and if there is an inaccurate situation during the learning period, it is necessary to redesign the network and learn again. Width learning is that after designing a good network, when faced with inaccurate learning, it can be expanded incrementally at any time in a horizontal way, that is, by adding neurons to improve accuracy. This incremental learning mode is also suitable for the real-time entry of data into the trained neural network model without having to retrain the whole collected data.

In the field of security, the application of wide learning network is mainly manifested in two aspects: one is to improve the reliability of artificial intelligence recognition. For example, in the face recognition algorithm training process, the best data is of course the clean face data of high-definition frontal face, but in fact, in the process of testing and reasoning, a lot of face data is not perfect. There will be occluded (sunglasses, mask), blurry, non-positive face angle of the face photos. In the process of algorithm training, we can learn the network architecture based on width, and even generate some defective image samples by merging clean face images and defective face images together, so as to improve the recognition accuracy of the algorithm to defect images, so as to improve the scene adaptability of the face recognition algorithm in complex scenes. The second is to solve the problem of data tagging, in the artificial intelligence algorithm training process, data tagging is also very important, if tagging wrong, then no matter how accurate the algorithm is, the training results will not be ideal. Through the algorithm model constructed by the width learning network, the problem of algorithm labeling error can be well solved.

Through a large number of tests by the research team, we can see that width learning (BLS) and its various variants and extended structures have good development potential, and show its excellent performance of high speed and high precision in practical applications. At present, width learning has been applied in many technical fields, such as time series, high spectral analysis, brain-computer signal analysis, fault tolerance, gene identification and disease detection, gait recognition, 3D printing and intelligent transportation. With the continuous deepening of artificial intelligence technology research, width learning, an efficient incremental learning system that does not need depth structure, is expected to accelerate the development of artificial intelligence.

Https://www.toutiao.com/a6763060586705060359/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.