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

Never waste a crisis.

-- Churchill

During the epidemic, Hangzhou's urban brain, health code, Ali whole gene sequencing, and CT images fully demonstrated China's urban governance capabilities with artificial intelligence and other new infrastructure.

The two sessions, which are under heated discussion, will take the construction of new infrastructure as the center, take the new development concept as the guide, and be driven by technological innovation to comprehensively develop new technologies such as 5G and artificial intelligence, so as to speed up the landing of industrialization.

Based on the brief deduction of the problems existing in the field of artificial intelligence, this paper briefly deduces the course of artificial intelligence after coming out of the cold winter.

I. what are the problems encountered in the production of raw materials (data) of artificial intelligence

Application status of data layer:

(1) unable to save-- the challenge of massive data storage under the tide of digitization

The sharp increase in the amount of data is an important reason why the data can not be saved. According to global statistics, the amount of data in enterprises is moving from PB to EB, and the amount of data will grow rapidly from 32.5ZB in 2018 to 180ZB in 2025.

As the storage system is still the traditional architecture and cost and other reasons, less than 2% of the current enterprise data is saved, the problem of "not being able to store" data is becoming more and more serious.

Lack of storage scalability: traditional storage consists of an independent controller and a hard disk frame, which can be cascaded with a new hard disk frame when the capacity is insufficient, but due to the limited processing capacity of the controller, the storage expansion capacity is very limited. The type of storage protocol is single: unstructured data has gradually become the main body of enterprise data. With the expansion of e-commerce and the Internet of things, 80% of the new data is made up of all kinds of audio and video, logs and other unstructured data. However, the type of traditional storage protocol is single, which can not meet the access needs of block, object, file, big data and other diverse data at the same time, so enterprises have to add a new storage device for each new data type, which increases the difficulty of efficient use of storage resources. Storage costs are still high: more and more companies are choosing to keep their data for a long time. Since 2017, mobile operators have increased the storage cycle of their device logs from 2 months to 6 months due to compliance requirements.

This means that the device size of its data storage server will be at least tripled. In the traditional architecture, the server continues to expand due to storage requirements, but the utilization rate of CPU is always in a low state, resources can not be used reasonably, which will undoubtedly cause greater pressure on procurement costs and maintenance costs. Companies have to give up a lot of valuable data because of the storage cost.

(2) the flow is immobile-- the long-standing problem of data isolated island.

The value of isolated data is not significant, and only when the data flows like water can we break down the "data barrier" and maximize its value.

"three types of isolated islands" of data:

Application island: the data generated by different applications are stored in different storage systems, and the data can not be shared with each other because of their own characteristics, that is, the problem of "application island" is formed. Management silos: in order to protect and use production data, a copy of production data is copied to various systems (such as backup, disaster recovery, archiving, development, testing and analysis systems) for management and use. Even the same data needs to be stored, managed and used separately in order to achieve different purposes, that is, the problem of "management island" will be formed; geographical island: due to the upgrading of enterprises, there will be multiple sets of storage devices, such as production environment, non-production environment, cloud environment and edge environment, enterprise data will be stored in different places, forming a "geographical island" problem.

(3) poor use-applications are complex and inefficient as a result of insufficient data supply

The huge amount of data has given birth to unprecedented opportunities as well as great challenges. Some people even say that there is never a shortage of data, and that more data becomes a burden. Others say that data is only a resource, not an asset, and it is difficult to generate value. The fundamental reason is that the data is not used well, and the data has no release value. The main reason that affects the release of data value is the insufficient supply of data, unable to feedback the nature of the business, and support business decisions: a large amount of data is not stored.

Enterprises produce a large amount of data every day, but the traditional data entry requires manual planning in advance, which leads to a large number of unstructured data and some new data can not enter the system (such as IoT data, video data, picture data, etc.). The lack of data will weaken the perception of the business and can not reflect the nature of the business in a true and timely manner.

Can not find data traditional enterprises usually use data tables to manage and analyze data, larger corporate data tables can even reach millions, and scattered in various business systems. Without a unified data catalog and global data view, finding specific data in millions of reports is like looking for a needle in a haystack, unable to cope with flexible business needs.

Next, committed to data fusion, explore the value of data, companies with data will continue to redefine the development of the artificial intelligence industry.

Second, the technology and application trend of artificial intelligence

From the technical level, most of the successful AI technologies in recent years are based on deep learning and neural networks, and the current hot research direction is still based on the improvement of these methods and the integration with other ideas such as adversarial learning or reinforcement learning.

(1) further breakthroughs in in-depth learning, generalization and interpretability.

In terms of applications, these technologies have entered our daily life, such as intelligent recommendation, machine translation, chat robots and various image beautification applications.

For the near future, people generally believe that there is room for further development of deep learning, and even hope to revolutionize our way of life. For example, many people believe that successful autopilot technology will completely change our lives and travel.

At the same time, due to the difficulties of deep learning in generalization and interpretability, some people think that deep learning will be difficult to play a really important value in some areas where there are strict requirements, and further breakthroughs need to explore a wider range of methods.

Model interpretability of deep learning has become a major obstacle to the wide application of deep learning, which has caused some discussion and controversy in legal and ethical aspects. It can be expected that technical exploration and social discussion in this area will continue.

The integration of deep neural networks and new ideas Al continues to enter human daily life, intelligent speakers enter the home, face recognition guards the streets, machine translation, self-driving cars have been tested on the road, and digital assistants are learning to deal with more and more complex tasks. Al apps will continue to enter more daily life scenarios and continue to change the way we live.

(2) efficient and low-cost models are still the focus of research.

Efficient and low-cost models although quite a few models have been able to run effectively on lower-cost hardware, in order to train enough practical models, current deep learning methods often require large-scale data sets and costly computing hardware.

How to design efficient training methods has become an important research topic at present. For some models that have been born, such as Mobile Net, Shutfte Net and Pelee Net, their practicability is still considered to be lacking. Efficient and low-cost models will continue to be an important future direction in the AI field.

(3) the integration of 5G and AI is another major trend.

In the technology maturity curve (The Hype Cycle), we can fully see that the technology of infrastructure construction such as 5G has been pushed to the forefront of the wind and waves. From the technology maturity curve, we can easily see that the development of 5G and RPA has become a priority, while the attention to machine learning is on the decline.

For a certain technology, the development of technology itself will decline year by year with the attention of the media, while to maintain the vitality of a technology itself, it is more technology + scene, but also the practicability and replicability of the technology itself.

Therefore, for the future of artificial intelligence, the integration of technology and the integration of technology and scene has become a top priority. The establishment of AI platform has won opportunities for more people, and also makes the rapid integration of technology possible.

There are all kinds of uncertainties for the future, but it doesn't matter to follow the trend. This is the only thing we can firmly believe in. For each of us, finding the trend and adapting to the trend may be the only thing we can do.

Http://www.woshipm.com/ai/3907556.html

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.