In-Memory Sensing and Computing for Artificial Intelligence Applications

Wednesday - May 10 2023, 07:54 UTC - 1 year ago

A team of scientists have developed a device that is able to sense, store, and process light-sensitive data and images using a two-terminal solution-processable MoS2 based metal–oxide–semiconductor (MOS) device. The device provides high temperature retention, low power consumption, and is capable of in-memory computing. It is suitable for revolutionary AI applications such as autonomous cars and robotics.

Artificial intelligence (AI) and the internet of things (IoT) have led to the rapid expansion of sensory nodes, which can produce an enormous volume of raw analog data that is converted into digital data, and then transmitted to other units to perform computational tasks. Nevertheless, the conventional von Neumann architecture which consists of discrete devices results in data access and data analysis delays with high power consumption. This can be troublesome for revolutionary applications with stringent delay and power requirements such as autonomous cars and robotics, among others.

In a new paper published in Light: Science & Applications, a team of scientists, led by Professor Nazek El-Atab from the Smart Advanced Memory devices and Applications (SAMA) laboratory, Electrical and Computer Engineering Program of King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia, and co-workers from Electrical Engineering and Computer Science, Khalifa University, Abu Dhabi, UAE, have developed a single sensing-storage-processing node using a two-terminal solution-processable MoS2 based metal–oxide–semiconductor (MOS) device.

The device embeds a light-sensitive 2D material-based charge-trapping layer that mimics the human visual system. More specifically, the same device is shown to be capable of optical data sensing, storage and processing. The study highlights how advanced technology-based in-memory sensing and computing can improve the response time, area, and energy efficiency, overcoming the delayed data access and hardware redundancy issues in conventional von Neumann architecture.

"When the device is exposed to light, it would directly store the wavelength and intensity of light within it, as opposed to traditional devices where a dedicated photosensor would detect the intensity/wavelength of light, then convert the data into the digital domain using an analog-to-digital converter, and then transfer the data to a separate memory for storage," said the scientists, summarizing the operational principle of their in-memory sensing devices.

"When operated as a traditional memory, the device showed a decent memory window of approximately 2.8 V with an operating voltage of +6/-6 V, high-temperature retention (100°C) for 10 years, and excellent endurance (106 cycles) without any deterioration. Interestingly, the device showed a larger shift in the memory window from 2.8 V to more than 6 V when the optical light of different wavelengths was stimulated for 2 s during the program operation. This confirms that the device is able to sense light and to store it directly within the same node," the scientists added.

"The effect of the number and duration of electrical and optical pulses on the memory window suggested that these devices can also mimic the perceptual learning of the human visual system. To confirm this, a convolutional neural network (CNN) was used to measure the device's optical sensing, storing and processing capabilities. The array simulation received optical images transmitted over the blue light wavelength and performed inference computation to process and recognize the images. The results shows that our devices are able to recognize the objects in the images with 91% accuracy," the scientists explained.

"The determined approach is pottentially applicable to long-term data storage and in-memory computing of images and videos, which could be used for artificial intelligence and cognitive based computing," the scientists concluded.