Say no to choice difficulties, AI can help you choose clothes!

Produced by: Science Popularization China

Author: Wu Xiaosong (Fujian Institute of Research on the Structure of Matter, Chinese Academy of Sciences)

Producer: China Science Expo

"Hey, AI, I have a date tonight, help me pick out an outfit!"

"Okay, Master!"

…

This scene can now not only appear in science fiction movies, but can also happen in reality!

On January 28, 2023, the team of Researcher Huang Weiguo from the Fujian Institute of Research on the Structure of Matter, Chinese Academy of Sciences, and the team of Professor Wang Zhongrui from the University of Hong Kong published an article introducing a new type of optoelectronic materials and devices that can be used to achieve the goal of using Al to replace the human brain to recognize different fashion clothing.

The first step of AI in choosing clothes: “Sensing and computing in one”

This technology is similar to a wearable artificial eye that can perceive and recognize various images. If this goal can be achieved, it will open up new prospects for the future development of artificial intelligence technology.

How do we achieve this? First, design the material with purpose.

The human retina can sense light signals, collect and pre-process various dynamic images, and facilitate the next step of visual cognition. In order for AI to have visual intelligence, similar functions need to be realized. However, the signal perception, storage and processing units of traditional silicon vision chips are independent of each other, and the data transmission and analog-to-digital conversion between them will generate a lot of energy consumption and limit the computing speed. Moreover, as Moore's Law slows down, this limitation will become more serious.

Therefore, exploring new technical routes, developing new optoelectronic materials and devices, and giving them the characteristics of "perception and computing integration" are of great significance for realizing edge computing devices with low power consumption and high computing speed.

Excellent semiconductor materials can have "memory"

To integrate perception and computing, we must first realize "perception", that is, transient memory behavior.

The researchers proposed a materials-algorithm co-design strategy to develop a semiconducting polymer ( p -NDI**) with efficient exciton separation and space charge transport properties .

Figure 1 Morphology analysis of p-NDI film

(Image source: References)

Normally, when light shines on a semiconductor, some excitons are generated in the semiconductor. An exciton is a complex of an electron and a hole (Figure 1i). The electrons generated after the decomposition of the excitons participate in charge transfer, thereby increasing the photocurrent.

However, due to the Coulomb force, electrons and holes will eventually bind together and annihilate, causing the number of electrons participating in the charge transport of the p-NDI semiconductor to gradually decrease, resulting in a slow decrease in the current (Figure 1k), forming a transient memory behavior.

Figure 2 Comparison of the photocurrent response of conventional semiconductors and p-NDI, and detailed semiconductor design principles of the sensor-side RC system

(Image source: References)

As can be seen from the above figure, after the light is removed, the current of the transistor made of common semiconductors immediately returns to the initial value or there is almost no photocurrent.

When the light is removed, the current of the transistor based on p-NDI decreases nonlinearly over time, forming a transient memory behavior, which is similar to a type of memory in the human brain. For example, when we see a certain piece of clothing, it stays in our brain for a short time and then slowly disappears. This process is transient memory.

These properties are crucial for optical signal preprocessing and for performing “reservoir computing” in optoelectronic sensors.

The researchers showed through experimental analysis that p-NDI has a relatively regular out-of-plane orientation and a poor in-plane orientation. Such a structure not only ensures effective spatial charge transport, but also controls the speed of electron and hole recombination in space (Figure 1i). All these characteristics contribute to the excellent light response behavior and transient memory characteristics of transistors based on p-NDI.

The birth of the vision chip

Figure 3 Schematic diagram of optical multi-task learning

(Image source: References)

Figure 4 Event-based video classification using the DVSGesture128 dataset

(Image source: References)

The researchers used the above semiconductor materials to prepare a "reservoir computing" neural vision chip with multi-task recognition capabilities (Figure 4). After several simulations, practices and verifications, the "reservoir computing" device has recognition rates of 98.04%, 88.18% and 91.76% for handwritten letters, numbers and clothing, respectively. In addition, the recognition rate for different dynamic gestures (waving left hand, waving right hand and clapping) is as high as 98.62% (Figure 4).

This technology not only overcomes the bottleneck of large energy consumption in traditional sensor-computing systems, but also provides a promising material-algorithm co-design strategy for wearable, affordable, and efficient photonic neuromorphic systems with multi-task learning capabilities.

So what is the connection between these technological designs and our real lives?

They look very "high-end", but in fact, they are inseparable from our daily life and make our lives more convenient. For example, for people with impaired vision, artificial vision chips can undoubtedly help them "see" the world again and get rid of the various troubles caused by impaired vision in life. The integrated vision chip can also use its powerful computing power to quickly judge the road conditions and provide more protection for unmanned driving. Of course, there are many other application scenarios...

summary

With the collaboration of materials and algorithms, in the not too distant future, we may really have an AI butler that can help us choose clothes. The thought of this is indeed a little exciting!

Note: The related research of this article has been published in Nature Communications, and the title of the paper is Wearable in-sensor reservoir computing using optoelectronic polymers with through-space charge-transport characteristics for multi-task learning.

Editor: Sun Chenyu