MEXT-Kakenhi Grant-in-Aid (S) "Innovative Edge AI: Building the Science of Information Processing for Long-Timescale Human Interaction" Apr. 2025 - Mar. 2030
Members
Our project is driven by a dynamic team of seven researchers, reminiscent of Akira Kurosawa's legendary film, Seven Samurai. As below, we've illustrated our members' primary contributions across areas such as neuromorphic devices, architecture, and algorithms. Additionally, our research targets include ASIC development for CMOS circuits, FPGA-based emulation, reservoir computing exploration, and attractor extraction with learning applications. However, these classifications are merely convenient labels. Over the past five and a half years, our team has engaged in a JST-CREST research project, consistently embracing interdisciplinary collaboration beyond these categories. We intend to continue this approach in the current project. Furthermore, we are excited to welcome two research collaborators from international institutions, as well as numerous postdoctoral researchers and graduate students, enriching our team's diversity and expertise.
Outline of the Research
Nowadays, data is transmitted to cloud-based supercomputers for deep learning, enabling advancements unimaginable a decade ago. However, edge devices handling personal biometric data face risks of hacking and privacy breaches (Fig. 1, top left). Eliminating internet connectivity is ideal, but deep learning requires high power. Using general-purpose data for training and limiting edge processing to inference is an option but lacks individual adaptability. Moreover, human-related phenomena occur on long time scales (Fig. 1, top right), making ultra-low-power machine learning difficult. To address this, we propose Slow Electronics, designed to process slow phenomena. The brain, an ultimate Slow Electronics system, operates on just 20W to process temporal data. One of its key models is Leaky Integrate-and-Fire (LIF) neurons (Fig. 1, bottom left). However, high-speed LIF neurons require batch processing, increasing power consumption (Fig. 1, bottom right). Implementing long time-constant LIF in analogue circuits also demands impractically large capacitors. This project develops neuromorphic devices using oxygen vacancies, proton drift, and CMOS-based voltage-controlled oscillators (VCO) (Fig. 2, left), alongside reservoir circuits and brain-inspired algorithms (Fig. 2, right).
Figure 1. This study addresses the societal demand for real-time processing of privacy-sensitive data on power-limited edge devices. However, many time-series data are inherently "slow," making conventional approaches inefficient. Overcoming this challenge requires a new type of electronics capable of efficiently handling slow phenomena with low power consumption.
Figure 2. We develop neuromorphic devices with long time constants, a reservoir circuit, and brain-inspired algorithms.
Innovative Edge AI = Slow Electronics
This study aims to establish the foundation of Slow Electronics, a complementary approach to modern digital electronics. Its key advantage is real-time interaction with the environment while continuously learning and processing information. Traditional digital electronics rely on predefined models, but as complexity increases, they face computational explosion (Fig. 3 top). Slow Electronics overcomes this by utilizing reservoir computing with slow neurons (time constant >100s). This method, leveraging attractor generation instead of physiological models, successfully predicted blood glucose fluctuations 30 minutes ahead (Fig. 3 bottom). By applying real-world time-series data, this research builds theoretical foundations from the bottom up, presenting a novel approach to efficient information processing.
Figure 3. Slow Electronics learns while interacting with the environment in real time. Predefining models for all events causes computational explosion, but we successfully predicted blood glucose without a model; theoretical framework of Slow Electronics.
Contacts
Isao H. Inoue (Project Leader).