For several decades, the rapid growth of the electronics industry has been driven by the continuous miniaturization of semiconductor devices. However, this trend is now approaching its physical limitations. Concurrently, the advancement of artificial intelligence (AI) is placing unprecedented demands on computing performance, leading to an exponential increase in energy consumption 

The progression of AI further exacerbates the memory wall issue, a critical bottleneck between processing and memory units. Additionally, the emergence of sensor-intensive applications, such as autonomous vehicles and humanoids, is introducing significant latency challenges between input (sensors) and computing units.

Addressing these challenges requires the development of novel computing architectures, including neuromorphic computing and in-sensor computing. Additionally, edge computing is required to facilitate energy-efficient signal processing at the individual edge device level.

In response to these demands, we aim to explore and develop new semiconductor devices tailored to these emerging computing paradigms, which are vital for the continued evolution and enhancement of AI technologies.

To implement artificial intelligence with performance comparable to or surpassing that of humans, the deployment of a large-scale artificial neural network is essential. However, utilizing current computing methods to train and infer such extensive neural networks demands an enormous amount of energy. Consequently, the development of computing systems that emulate the energy efficiency of biological systems is imperative.

Biological sensory systems excel in processing, from external stimulus recognition to inference, with remarkable energy efficiency. These systems preprocess sensory signals from sensory organs and nerves, transmitting only meaningful signals to the brain. This selective transmission allows the brain to infer and perceive efficiently. Despite this, much of the current research is concentrated solely on neuromorphic computing, which emulates the brain's efficiency.

Our approach extends beyond this focus. We study and develop an artificial sensory system that effectively mimics the entire sensory neural system, encompassing sensory receptors, sensory nerves, and the sensory cortex. By doing so, we aim to create a more comprehensive and energy-efficient AI system.

Existing computing architectures are susceptible to significant energy consumption and latency issues. By integrating and leveraging in-sensor computing technology, which preprocesses data at the sensor level, along with neuromorphic computing technology, it is possible to overcome the limitations of traditional computing structures. Furthermore, configuring a computing platform using three-dimensional integration on a single substrate can achieve high energy efficiency and address latency concerns.

Building on technological advancements achieved through horizontal integration on a single silicon layer, our research focuses on developing a three-dimensionally integrated system to enhance performance. Utilizing a silicon substrate, we design CMOS-based circuits and memories on the silicon layer, while sensors capable of detecting various stimuli are positioned on the upper layer. This approach enables the creation of an AI system that can efficiently perform stimulus detection, data storage, and processing, all in a highly energy-efficient manner.

Building upon the development of this system, we are advancing on-device AI. For on-device AI to be effectively applied in autonomous vehicles and robots, innovation at the software level alone is insufficient; hardware innovation is equally essential. Our research aims to lead this critical innovation.