The world's first on-chip learning memristor storage and calculation integrated chip developed by Tsinghua University was published in Science magazine

This type of memristor storage and calculation integrated chip has positive significance for overcoming the "stuck neck" key core technology.

Tsinghua University’s official Weibo released an important result on October 9. The school successfully developed the world’s first memristor storage and calculation integrated chip that supports on-chip learning.

Recently, Professor Wu Huaqiang and Associate Professor Gao Bin of Tsinghua University have made major breakthroughs in the field of memristor storage and calculation integrated chips. Based on the integrated storage and calculation computing paradigm, they successfully developed a chip that supports on-chip learning. This research result has been published in the latest issue of the international scientific journal "Science"

According to Tsinghua University, the memory resistor (Memristor) is the fourth type of circuit after resistors, capacitors, and inductors. Basic components. It can still "memorize" the passing charge after the power is turned off, so it can become a new type of nanoelectronic synaptic device

Since 2012, the teams of Qian He and Wu Huaqiang of Tsinghua University have been working on Collaborative research on memristive devices, prototype chips and system integration has gradually solved the problem of AI computing power bottleneck. Their new research has overcome the "stuck neck" problem in key core technologies to a certain extent

The paper "Edge learning using a fully integrated neuro-inspired memristor chip" is as follows.

Please click the following link to view the paper: https://www.science.org/doi/full/10.1126/science.ade3483

A research overview refers to a general understanding and overview of a certain research field. It usually includes a brief introduction to the field's historical background, research objectives, methods, and results. The purpose of the research overview is to help readers quickly understand the basic situation in this field and provide a basis for further in-depth research. A research overview is usually a part of a research paper, report, or academic article that helps readers gain an overall understanding of the research content. By reading the research overview, readers can learn about the important research progress and existing knowledge gaps in this field, thereby providing reference and inspiration for their own research

We know that memristor-based computing technology This technology has received considerable attention recently for its potential to overcome the so-called "von Neumann bottleneck" of traditional computing architectures. What’s special about memristors is that they can enable real-time, energy-efficient on-chip learning for a variety of edge intelligence applications, even though the implementation of fully on-chip learning is still challenging.

A schematic diagram of edge learning using neurally inspired memristor chips is shown below. Figure 1 illustrates the human brain’s ability to improve learning. Figure 2 shows the design and future applications of a memristor-based neuro-inspired computing chip. This chip is designed for complete on-chip learning, integrating all necessary modules with the memristor array, so that edge AI devices have learning capabilities and can quickly adapt to new scenarios

In order to solve related problems, doctoral student Zhang Wenbin and postdoctoral fellow Yao Peng from the School of Integrated Circuits of Tsinghua University proposed a solution called Memristor Characteristic Symbol and Threshold-Based Learning Architecture (STELLAR) and successfully produced A full system integrated chip. The chip includes multiple memristor arrays and all necessary complementary metal oxide semiconductor peripheral circuits required to support complete on-chip learning

Figure 2 below shows the memristor feature architecture design for on-chip learning, A is STELLAR architecture used in memristor chips, B and C are comparisons of classification accuracy, D is the weight with differential conductance pairs (left) and 1T1R (middle) and 2T2R (right) configurations, E is the cyclic parallel conductance adjustment scheme .

Figure 3 below shows the memristor chip used for on-chip learning. A is the architecture overview, B is the optical microscope image of the chip, and C is the cross-sectional transmission of the 2T2R cell. Electron microscope image.

Researchers demonstrated end-to-end on-chip improved learning on a variety of tasks, such as motion control, image classification, and speech recognition, achieving software-like accuracy and Lower hardware costs. This work marks an important step in the field of in-memory computing.

Figure 4 below shows an example of improved learning using a memristor chip. A shows the motion control task and its control system, B shows the learning of new samples of light chasing cars, and F shows the learning of new categories in the image classification task

Let’s take a look at the following animation demonstrations.

First, we will discuss a new category learning task for handwritten digits

In addition, learning can be improved in the field of motor control. As shown below, prior to improved learning, the forward-moving blue car tended to miss the target red car.

After learning and improving, the blue car moving forward will first make a backward movement to adjust, and finally continue to move forward towards the target red car

Not only that, before learning is improved in bright scenes, the blue car will often deviate from the target red car during the following process.

After improved learning in bright scenes, the blue car adapts well and always follows the target red car.

As the co-first authors of academic papers, Zhang Wenbin and Yao Peng were exposed to a large number of scientific research in different directions such as semiconductors, microelectronics, software algorithms, and brain-inspired computing during their Ph.D. knowledge, and has accumulated rich research and development results and rich engineering construction experience.

##.

Reference report:

The content that needs to be rewritten is: https:// mp.weixin.qq.com/s/w0VZNIQ1KbClJJ8c05hPqg

###

The above is the detailed content of The world's first on-chip learning memristor storage and calculation integrated chip developed by Tsinghua University was published in Science magazine. For more information, please follow other related articles on the PHP Chinese website!