The needs of the industry determine that the artificial intelligence tasks to be completed are becoming more and more complex. Lightweight artificial intelligence must accelerate the calculation efficiency and increase the calculation density to achieve the ultimate efficiency.
——Leng Cong, associate researcher at the Institute of Automation, Chinese Academy of Sciences
◎Our reporter Lu Chengkuan
The complexity of artificial intelligence algorithms has risen sharply, the cost of energy consumption for neural network calculations has become higher and higher, and the flood of data has accumulated dams...In recent years, the development of artificial intelligence has encountered more and more bottlenecks.
How to front-end and lighten the artificial intelligence model and its computing carrier has become an urgent problem to be solved.
The newly emerging lightweight artificial intelligence has high hopes, and the second half of the artificial intelligence competition with "lightweight" as the match point has arrived.
To this end, on April 27, a reporter from Science and Technology Daily interviewed relevant experts from the Institute of Automation, Chinese Academy of Sciences.
Lightweight becomes artificial intelligence second half match point
In industry applications, artificial intelligence technology mostly relies on massive training data and computing power support from large-scale servers.
However, in recent years, with the gradual slowdown of Moore's Law in the field of information technology, the development of hardware has become increasingly difficult to meet the trillion-scale storage and computing power requirements of current artificial intelligence models, data barriers, storage skyrocketing, and privacy leaks , High energy consumption and other issues follow.
"At present, the demand for rapid response, privacy protection, energy saving and emission reduction of artificial intelligence equipment and applications is becoming more and more prominent. Lightweight artificial intelligence has emerged as the times require, and high hopes are placed. In 2020, the "MIT Technology Review" will be lightweight. Artificial intelligence is listed as'the world's top ten breakthrough technologies'." said Cheng Jian, a researcher at the Institute of Automation, Chinese Academy of Sciences.
The so-called lightweight artificial intelligence refers to the use of a series of lightweight technologies to improve the efficiency of chips, platforms and algorithms, and to achieve low-power artificial intelligence training and application deployment in a tighter physical space, without relying on the cloud. Artificial intelligence that can realize intelligent operation through interaction.
Lightweight artificial intelligence was rated as one of the “Top Ten Global Breakthrough Technologies”. The reason given by the MIT Technology Review is that lightweight intelligence makes existing services, such as voice assistants and mobile phone camera, even better. Faster, you don’t need to connect to the cloud every time to run deep learning models; in addition, lightweight artificial intelligence will also make new applications possible, such as mobile-based medical detection and analysis, and autonomous vehicles that require faster response time. In addition, localized artificial intelligence is more conducive to privacy protection, and user data no longer needs to leave the device to realize the evolution of service functions.
"More importantly, lightweight artificial intelligence pushes artificial intelligence to a more mainstream. It greatly reduces the difficulty and cost of deployment of artificial intelligence systems, turning artificial intelligence from a high-threshold technology giant competition to easier to benefit the people. The intelligent ecology of life." Cheng Jian said that in the field of artificial intelligence, the second half of the match point with lightweight has arrived.
Extreme efficiency and low energy consumption are the ultimate pursuit
In terms of performance, lightweight artificial intelligence is doing subtraction, reducing energy consumption, lowering the requirements for hardware platform performance indicators, and reducing the need for communication with the cloud.
However, "in essence, the lightweight core is doing addition." said Leng Cong, an associate researcher at the Institute of Automation of the Chinese Academy of Sciences. The ultimate efficiency can only be achieved by increasing the calculation speed and calculation density.
In Cheng Jian's view, it is a challenging task to reduce the weight of the artificial intelligence model and its computing carrier under the premise that the accuracy is close to lossless.
To solve this problem, it is necessary to carry out lightweight design, calculation acceleration, and design of new computing architectures for the neural network to realize the hardwareization of the model. This requires starting from both software and hardware.
On the software, carry out model and algorithm innovation, through lightweight model design, matrix decomposition, sparse representation, and quantitative calculation to achieve model miniaturization and calculation acceleration; on the hardware, it must be through pipeline design, storage mode design and other means Innovate the hardware architecture, and realize the lightweight of artificial intelligence through the collaborative design and optimization of software and hardware.
"Although the neural network calculation is performed by hardware, the neural network structure and artificial intelligence platform determine the amount of calculation and the calculation method." Leng Cong said frankly, so the ultimate lightweight must be the collaborative lightweight of software and hardware-based on Complex artificial intelligence application scenarios fully combine chips, platforms and algorithms for joint acceleration.
As the hardware carrier of artificial intelligence, artificial intelligence chips must achieve higher performance, higher efficiency, lower power consumption and smaller size.
Only in this way can there be a sufficiently affordable and efficient computing platform to meet the needs of the industry, to carry complex artificial intelligence tasks, and to migrate reasoning and computing from the cloud to the terminal.
At the same time, the lightweight artificial intelligence platform should train and run artificial intelligence algorithms with lower power consumption to maximize the ability of the hardware to be explored.
More importantly, the neural network model using lightweight technology should be small in scale, less computationally expensive, and maintain good accuracy.
Lightweight artificial intelligence in the future will empower everything
Cheng Jian introduced that the Institute of Automation of the Chinese Academy of Sciences is a pioneer in lightweight artificial intelligence. It has started the research on soft and hard collaborative lightweight technology very early and is in the forefront of the world.
As early as 2016, when the convolutional neural network was applied on a large scale, the Institute of Automation of the Chinese Academy of Sciences published many important papers in the field of lightweight neural network models in the top international artificial intelligence journals, becoming the first artificial intelligence in the world. One of the light-weight research institutions, and the related results have attracted wide attention from many experts at home and abroad.
"The lightweight artificial intelligence platform QEngine and lightweight algorithms we designed and developed have been deployed on hundreds of thousands of terminals. In 2019, in the micro-network challenge competition at the International Neural Information Processing System Conference, we competed with ARM, IBM, Qualcomm, and Xilinx. Other international first-class chip companies competed in the same field and won the double championship in the image category of lightweight neural network architecture." Cheng Jian said.
In 2020, the world’s first ultra-low-bit quantitative neural processing chip (QNPU) independently developed by the Institute of Automation of the Chinese Academy of Sciences was successfully taped out, bypassing the “memory wall” problem that has attracted much attention in the field of chip computing. , Computing structure, edge computing and other aspects to achieve revolutionary changes.
"The launch of the chip also marks that the Institute of Automation has become one of the few institutions in the world that has a full-stack lightweight artificial intelligence technology of'artificial intelligence chip-platform-algorithm'." Leng Cong said.
In the future, more and more miniaturized devices driven by artificial intelligence will appear around us.
Lightweight artificial intelligence terminals composed of artificial intelligence chips, platforms and algorithms will be used in more and more scenarios.
“For example, in the power industry, my country’s power transmission lines cover a wide range, the natural environment in the wild is complex, and the maintenance and repair operations are dangerous and difficult. We designed autonomous inspection drones, defect recognition and analysis portable terminals, and channel visualization intelligent sensing cameras. With a variety of intelligent identification, detection and analysis functions, it can ensure the safety of transmission and distribution lines and the stability of the power system." Cheng Jian gave an example.
At the same time, in the consumer electronics industry, lightweight algorithms and lightweight neural network computing architectures designed by automation such as dark light enhancement and super-resolution also provide image enhancement effects for mobile phone terminals and security terminals.
Cheng Jian said that lightweight artificial intelligence will empower everything in the future, so that every device has the ability to perceive the environment, human-computer interaction, and decision-making control.