Tian Yulong, chief engineer and spokesman of the Ministry of Industry and Information Technology (MIIT), said at a press conference held by The State Council Information Office on Sept 13 that the industrial chain layout of emerging industries, such as artificial intelligence and quantum communication, should be strengthened. In this regard, some policies have been introduced to guide and support local governments, combine their own characteristics and regional strategies, enable them to form industrial chains with their own characteristics, and promote the clustering of emerging industries.
"Ai has greatly changed the way humans live and produce." Lenovo innovation technology conference in 2021, lenovo group senior vice President and chief technology officer RuiYong analysis said that lenovo is assigned to the industry practice, found that in all walks of life in the process of the intelligent transformation are faced with a series of common pain points, can be divided into three categories, one is the model building, the second is less data samples, slow scene adapter is three.
Rui Yong observed that in enterprises, many business experts are familiar with specific scenarios, but they may not be ai experts and do not know how to use AI to build appropriate models. The traditional approach is to ask artificial intelligence experts to build models manually, which is not only time-consuming and labor-intensive, but also difficult to make full use of computing resources to achieve the joint optimization of accuracy and delay. Another approach is to use machine learning technology to search for neural network architecture, which can save labor costs but consumes a lot of GPU computing power.
"When the model is deployed on the side or side, it is difficult to collect valid sample data for new problems and tasks." RuiYong said.
In Rui Yong's opinion, when a new scene needs to be built, such as a new intelligent production line or an intelligent factory, how to make use of existing capabilities to quickly replicate a complete set of solutions on a large scale? This requires consideration of a number of factors, such as which ai capability components will be invoked by the new task scenario? In what form do they combine and connect to work? Which devices should they run on to meet the requirements of computing power, storage, network, power consumption and other resources? Taking these factors into account is difficult and often requires years of experience and trial and error to find a better solution. This leads to the third pain point, how to adapt quickly and efficiently to new situations.
In view of the difficulty in model construction, Rui Yong said that Lenovo proposed the "multi-level fine-grained neural network architecture search" technology, which can save time and efficiently build high-precision models for business personnel. First of all, in order to reduce the search space, Lenovo proposed an incremental search method between modules, between layers and within layers. Compared with the traditional neural network architecture search method, a network architecture with optimal accuracy and minimum delay can be obtained faster. On this basis, the network weight preheating technique is used to assign priori initial values to the weights of the network architecture through nearest neighbor sampling, so that the model training can converge quickly. Taking the image classification problem as an example, with the help of this technology, the GPU computing power required to build a model with the same accuracy is only 1/10 of the common methods in the industry.
In order to solve the pain point of the industry transformation with less sample data, Lenovo combines "data enhancement" and "meta-learning" to create a "small sample lifelong learning technology". Rui Yong gave an easy-to-understand example -- learning to drive. "For example, you know how to drive a small car. Now you need to learn how to drive a big truck. There are a lot of differences, but if you can drive a small car, it's easier to learn to drive a big truck than to start from scratch, because there are some foundations to build on. That's how meta-learning works. By drawing on similarities between different tasks, we can achieve task-level generalization of learning ability and improve the model's ability to constantly adapt to new tasks." RuiYong said.
For scenario adapter (slow industry intelligent transformation, RuiYong says, had the lenovo solution automatically generated based on multi-objective optimization technology, users only need to input the new scene tasks and constraints, algorithm can automatically determine which model need, to run on any device can achieve the desired performance, how to collaborate between model and so on, Thus the optimal complete scheme is constructed. With this technology, it takes weeks to build an intelligent scheme for a new scenario, but now it can be done in hours.
These schemes to help the intelligent transformation of the industry and solve the pain points are mainly supported by lenovo Brain, an artificial intelligence platform for the intelligent industry. "Lenovo Brain's technologies are enabling the intelligent transformation of many industries." RuiYong said.
Manly Battery which located in Shenzhen,China. A leading battery manufacturer mainly offering rechargeable Lithium batteries,any query pls feel free to contact info@manlybatteries.com.
