It is our tremendous pleasure to welcome Professor Anind Dey of the University of Washington to SFC as a Guest Professor (International). Prof. Dey is Tadashi Okoshi’ PhD co-advisor in the Keio/CMU collaboration, one of the originators of context-aware computing, and a leading authority on technologies for recognizing human behavior and states from large-scale data. As the main event of this week, Anind gave a special lecture during our research group time on the recognition of “routine behavior” as a key to understanding human behavior through data and AI.
Watching our students from freshmen all the way to PhD candidates actively raising questions and discussions was quite an impressive moment, feeling like “witnessing” the breath of the next generation globally-minded talent emerging from SFC.
With Prof. JeongGil Ko, Prof. Chenren Xu, and Prof. Dey visiting SFC one after another from Korea, China, and US through Keio’s Global Faculty Program, we can feel the baseline level of international awareness among our students in the lab steadily rising, becoming something they simply take for granted.
More plans are underway with each of these professors for international co-authored papers, possible internships, and other forms of academic exchange.
we am genuinely excited about what lies ahead, and at the same time, we feel deep gratitude toward these professors who kindly accepted my request and traveled all the way to SFC.
卒業生 大崎君の研究「Overcoming eBPF Constraints: Towards Practical ML/NN-based Anomaly Detection」が情報処理学会論文誌に採択されました．学部4年次の成果です．
概要:
Machine learning (ML) and neural networks (NN) in extended Berkeley Packet Filter (eBPF) and eXpress Data Path (XDP) are being explored for network anomaly detection to balance accuracy and throughput. However, eBPF’s constraints, such as the instruction count limitation and the absence of floating-point support, make the implementation of such algorithms challenging. While previous works use workarounds, they lack cost analysis of the workarounds and performance comparisons with the user space approach. This paper investigates the practicality of overcoming these constraints in the context of ML/NN-based anomaly detection through evaluations of the cost and the impact of workarounds. We show that while instruction count is not a major concern, arithmetic errors in static fixed-point calculations negatively affect feature extraction and detection accuracy. However, we also show that dynamic fixed-point can match the detection accuracy in kernel space with that in user space. Furthermore, we demonstrate that the performance gap between eBPF/XDP and faster kernel-bypass technologies narrows to just 7% in ML/NN-based anomaly detection contexts. We conclude that ML/NN-based anomaly detection in eBPF/XDP is not only feasible but beneficial for real-world deployments compared to alternatives.
博士課程3年 濱中君の研究「Sensor-Augmented Voice Activity Projection for Enhancing Turn-Taking Prediction」が国際会議 SIGDIAL2026にacceptされました．本研究はNTTコミュニケーション科学基礎研究所との共同研究です．
Abstract:
Voice Activity Projection (VAP) has been actively studied to enable natural turn-taking in spoken dialogue systems, relying primarily on acoustic features. Visual cues such as head movements are also known to contribute to turn-taking prediction; however, camera-based approaches are affected by placement and lighting conditions and are not always reliably available to dialogue systems. As a camera-independent approach for directly capturing head motion, earable devices offer a promising solution. In this study, we propose Sensor-Augmented VAP, a framework that integrates in-ear inertial measurement unit (IMU) signals with a pre-trained VAP model via a lightweight residual fusion module. To validate our proposed method, we collected a dataset pairing conversational audio with in-ear IMU data, comprising 12 dyadic Japanese dialogues recorded using microphones and earbuds. Experiments in speaker-independent and speaker-dependent settings demonstrate that IMU fusion consistently improves weighted F1 score for shift detection and reduces VAP loss over the audio-only baseline. These results confirm that head-motion cues are effective for enhancing turn-taking prediction.
Abstract: Timber dwellings of the Miao people in Southwest China embody orally transmitted construction intelligence developed by Zhangmoshi master carpenters. While photogrammetry and laser scanning can capture geometry, they reveal little of the proportional and procedural logic guiding construction. This study proposes a rule-based parametric framework that encodes key Miao principles—Jian-xia (bay layout), Gua-zhu (bracket–column relation), and Bushui (roof slope)—using Grasshopper and Python. The system integrates terrain optimisation via Galapagos and is structured for interoperability with Heritage BIM environments. Validation on a representative Miao dwelling demonstrates geometric fidelity and adaptability. Through modular rule extraction and parametric recombination, the framework enables systematic variation while preserving vernacular coherence. The resulting system supports reconstruction, analytical comparison, and design exploration across different site conditions and parameter settings. By translating oral craft knowledge into computational rules, the research advances digital heritage and contributes to broader discussions of cultural algorithmicity and computational heritage in vernacular architecture. The approach further suggests a transferable methodology for encoding other tectonic timber traditions facing challenges of intangible knowledge loss.
Wang, Liuhan, Jin Nakazawa, and Hiroto Kobayashi. “Encoding Vernacular Intelligence: Parametric Rules for the Preservation and Fabrication of Miao Timber Architecture.” (2026).
Abstract: Edge computing plays a key role in smart-city deployments. Deploying AI models on edge devices reduces latency and privacy concerns. This work proposes QPilot, a reinforcement-learning-based adaptive model-switching method that dynamically selects detection models of different sizes for multi-object tracking (MOT) inference to balance efficiency and accuracy according to device compute capacity and scene complexity. In object-sparse scenarios, it improves tracking accuracy with only a small additional computational cost and achieves an optimal efficiency–accuracy trade-off in MOT. We validated the method on multiple hardware platforms, showing strong cross-device adaptability as well as feasibility and promise for edge environments.
