Shulou(Shulou.com)06/03 Report--

2020-04-10 21:16:01 author | Jia Wei

Proofreading | Yamashiguang

Edit | end of cluster

Once known as the troika of artificial intelligence-algorithms, computing power, data-it is time to reflect. In particular, the follow-up development of deep learning in the academic frontier is weak, and it may be difficult to support the continued upgrading of AI capabilities.

After entering the new decade, AI may have reached a turning point.

So how will AI technology be developed and applied next? With the end of the domestic epidemic, this has gradually become a key question that many researchers and practitioners in the field urgently need to answer.

On April 9, the academic Department of the Society of the China Association for Science and Technology, the China Science News, together with Tencent Association of Science and Technology and Tencent Development Research Office, held a very timely online forum with the theme of "artificial Intelligence: a New engine for the Integration of Science and Technology and economy". Many scholars and experts have answered many questions about the integrated development of artificial intelligence technology and economy after the epidemic.

AI Science and Technology Review focuses on artificial intelligence technology and selects the contents of the report by Shan Shiguang, a researcher at the Institute of Computing of the Chinese Academy of Sciences and the founder of China Science Shituo, to collate and share with you.

Researcher Yamashiguang's report is divided into two parts. First, from the perspective of research, he believes that AI methodology will change from "data-driven" to "knowledge + data joint drive"; in the latter part, he puts forward five views and suggestions from the level of industry development. These contents have profound insights.

1 from the perspective of academic frontier

Deep learning has become the end of stress.

In the past decade, the research of artificial intelligence has gradually transitioned from the methodology of designing artificial intelligence algorithms based on rules and knowledge to the methodology with data as the main driving force.

Under the guidance of this methodology, the current AI capability is supported by the "troika"-algorithm, big data, and strong computing power. The algorithm is mainly based on deep learning, and the data emphasize that it should be large enough (and supervised labeled data). Because deep learning often needs to set hundreds of millions of parameters through training, it also leads to a great reliance on the support of strong computing power.

Therefore, as long as the AI task meets the following two conditions, it can be solved:

1) dedicated AI tasks (rather than generic AI). For example, in medical images, the AI of pneumonia can only see pneumonia, not hepatitis, while the AI of CT can only see CT, not MRI.

2) "good" data is fertile (a large number of data are obtained through business models). The so-called good data, on the one hand, should have a larger scale, on the other hand, it should be well marked.

But such an AI is far from what we expected.

Dr. Hong Xiaowen, Director of Microsoft Research Asia, has proposed the following AI capability pyramid (yellow font is Yamashiguang added content):

According to this pyramid level, the current AI technology still stays at the second level of "perception and simple reasoning", which is equivalent to the level of non-human primates.

How to go one step further and achieve cognition, emotion, creation, and even wisdom?

Shan Shiguang believes that we need to improve at least the following abilities--

However, the existing AI methodology is not sufficient to support the continued upgrading of AI capabilities.

First of all, the algorithm, computing power, data these three carriages have been slightly weak. Both the improvement of computing power and the collection of big data need to invest resources, but most of the algorithms are one model for one task. Now there are many problems, but the methods are limited.

Secondly, deep learning still relies on big data and big arithmetic to carry out applied research in industry. due to the lack of strong theoretical support, some math power and data will be wasted on trial and error, which has become more and more obvious.

On the other hand, on the academic frontier, the potential of deep learning is limited. Why? Because it can not overcome the problem of high data dependence, it can not learn accurately and robustly based on "weak" and "small" data like human. The existing research also hopes to solve the problem of data dependence based on deep learning, hoping to give machine learning the ability to learn from weak supervision and small data, but in principle, if there is no other data or knowledge to support, it's almost impossible. We have to find a whole new way.

2 AI methodology in the next 10 years:

Knowledge + data joint drive

The upgrading of AI capabilities must rely on the upgrading of AI methodology, which boils down to the essence, or the innovation of algorithms.

Researcher Yamashiguang proposed the following AI algorithms that need to be developed urgently.

For example, can the algorithm reduce the amount of data to 10% or even 1% while maintaining the same ability as the original? In the existing algorithm (pure data-driven), can knowledge or other existing data or models be added to jointly drive knowledge and data? In fact, many research units have been doing it, and it is also one of the most worthwhile research directions in AI algorithm.

In view of the above, Yamashiguang made a detailed explanation from the perspective of "data dependence". As mentioned earlier, the current method of artificial intelligence is strong (strong supervision, large-scale) data dependence. But we know that this is not the case with human intelligence. Here are a few examples of human abilities:

Inductive and deductive reasoning: from individual to general, and then from general to individual

Analogy: similar deduction & transfer Learning

Learn a lesson: learn from a small number of mistakes (modify the model)

Prediction test: always predict and correct errors; self-correcting learning

Meta-method: Tao gives birth to one, one to two, two to three, and three to all things.

Integration: multimodal and multidisciplinary knowledge verification and fusion

Imagination and creativity: create something out of nothing, extrapolate rather than interpolate.

From these abilities, we can see that human intelligence is essentially a knowledge + weak (weak supervision, small sample) data-driven method. This characteristic is worthy of our reference.

Researcher Yamashiguang believes that human beings are able to learn small data because of the accumulation of knowledge. Therefore, how to integrate knowledge into the machine is very important for AI algorithm. The knowledge here may be either the knowledge summed up by human beings or the knowledge already learned by AI, which he calls "machine knowledge".

The so-called "machine knowledge", which is different from "human knowledge", is probably not readable or even understandable by human beings. As researcher Shan Shiguang said, when we have solved N tasks with algorithms (such as face recognition, monkey face, horse face, dog face, cow face, etc.), it is possible for machines to sum up some general task rules from these tasks. as a "meta" model.

With "machine knowledge", even small data / no data tasks can achieve good performance. For example, with the help of the above facial recognition models learned from people / monkeys, horses, dogs, cattle, etc., a "meta" model of faces can be obtained, which can be used to recognize koala faces, fish faces, panda faces, etc.

(now the method is often, to identify an animal, it is necessary to collect a large number of pictures of the animal's face, and to change the animal, it has to recollect and learn, which is neither efficient, nor elegant, or even ridiculously clumsy. )

In fact, this kind of research is essentially a problem of multitasking, which has been done in 2018, and is represented by the best paper by CVPR 2018, which studies the relationship between 26 different tasks and how they can support each other, thereby reducing the need for tagged data.

Based on the above analysis, Yamashiguang researchers believe that the methodology of AI has gradually become strong and big data driven in the past decade, but in the next decade or more, knowledge and data jointly driven will become the mainstream. Of course, the knowledge here not only refers to the knowledge that human beings can understand, but also may be the "machine knowledge" that many human beings cannot understand.

3. Five viewpoints & suggestions

In the second half of the report, researcher Yamashiguang put forward five major views and suggestions on the current development of artificial intelligence in China, as follows:

1. There is not a big gap in AI application research.

There is not much difference in applied research between China and the United States (Europe), but we still need to work hard on in-depth application in various industries (that is, in other research areas). For example, the application of AI in the field of bioinformatics has accumulated a lot in the West, and deep learning has also penetrated a lot, but our country needs to make efforts to catch up with and surpass in this area.

2. The gap in basic research of AI should not be underestimated.

The gap between China and the United States (Europe) is narrowing, but China's acceleration is not enough. The most representative method in the past decade has mainly come from European and American universities or companies, and it may take at least 5-10 years or more to be on an equal footing. The problem we have is the lack of understanding of the long-term nature of basic research and the lack of basic patience. The evaluation cycle of major basic research projects is too short, and even requires a clear technical route and results in two years. But basic research is often "inadvertently planted willow shade, willing to plant flowers do not blossom". Take the deep convolution neural network (DCNN) as an example, which is a product of the 1980s and did not begin to exert its power until nearly 20 years later.

3. There is a big gap in AI infrastructure.

This includes three aspects, namely, hardware, software and smart parts.

Domestic investment in the basic hardware platform is very large, but there is repeated construction, has not yet formed a joint force, can not give full play to its effectiveness; in addition, due to the weak theoretical basis of deep learning, invalid trial and error leads to the phenomenon of waste of arithmetic power.

On the other hand, China's investment in basic software platforms needs to be improved (at least 4-8 years behind North America), and the mainstream underlying framework for deep learning (TensorFlow,Pytorch,MxNet) is mainly built by North American countries. 10 years later, this aspect may become our "hidden danger" just like today's "chip industry". Recently, many enterprises in China (Baidu, Huawei, Kuangshi, Tsinghua, Pengcheng Lab.) have released or intend to release open source frameworks, hoping to form a joint force. Yamashiguang suggested that for low-threshold AI research and development platforms and tools, we need to seize the opportunity as soon as possible.

In addition, Yamashiko suggests that we should increase investment in the basic smart component system (including basic algorithm research), and suggest strengthening the research on the basic intelligent component system: starting from the existing computing centers and data centers in the past, how to build algorithm centers, knowledge centers, etc., it is suggested to build a national "knowledge center" as soon as possible. Including human knowledge center (general knowledge + domain knowledge) and machine knowledge (mature AI algorithm and model) center, so that the realized AI algorithm can be used like water, electricity and coal, so as to avoid a lot of repetitive work.

4. The tilt of AI personnel training is not enough.

Although it has been said recently that there is a large talent gap in AI, in fact, the number of graduate students majoring in AI is still insufficient. This deficiency is now mainly alleviated by other non-AI majors actively or passively AI, which is not a long-term solution. If the country really thinks that artificial intelligence is a strategic direction, it should give more preference to the training of AI talents, such as allocating more postgraduate places to AI majors.

5. The respective positioning of industry, university and research still needs to be optimized.

In recent years, we will notice a phenomenon, that is, the industrialization of enterprises, universities and institutions. Typically, companies are publishing articles, while university research institutions are doing short-term technology. There are many reasons, including how to evaluate the basic research of Long-term is worth discussing, hat culture has also given birth to a lot of fast-food scientific research results and so on. Recently, the state has repeatedly proposed to break the "four only", but it is even more suggested that we should not adopt a "one size fits all" scientific research evaluation model, but should adopt different evaluation methods according to different fields and different types of talents. In addition, the support orientation for enterprise scientific and technological innovation is also open to question. What kind of innovation enterprises should do and what kind of innovation universities and scientific research institutions should do may have a more correct classification.

Https://www.toutiao.com/i6814071107889398286/

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.