Dengzhe Hou (侯登哲)

I am an Assistant Professor at the Graduate School of Information Sciences (GSIS), Tohoku University, Japan, affiliated with the Yamada Laboratory and the International Liaison Office (ILO), with a concurrent appointment in the Social Integration Research Division of the Unprecedented-scale Data Analytics Center (UDAC).

My research asks when a measurement of a complex system can be trusted, whether that system is a brain or a model. I work on the reliability of EEG decoding, on cognitive probes for large language models, and on the neural mechanisms of voluntary, self-initiated attention. I received my Ph.D. under Prof. Satoshi Shioiri at the Visual Cognition and Systems Laboratory, and was a Visiting PhD Scholar in the Sydney Cash Lab at Harvard Medical School / Massachusetts General Hospital.

Research Interests

• Reliability of EEG decoding, including preprocessing-induced instability, per-trial uncertainty, and feature attribution
• Cognitive probes for AI systems, including working memory and cumulative state tracking in large language models
• Evaluation protocols that isolate what a model can actually do
• Neural mechanisms of voluntary, self-initiated attention, using EEG with simultaneous eye movements and gaze-contingent paradigms

My current work extends this toward aligning brain signals (EEG, fMRI) with large-scale AI models such as LLMs and vision-language models, comparing their internal representations.

Selected Research

Same Brain, Different Prediction

Across six datasets and four paradigms, up to 42% of trial-level EEG predictions flip when only the preprocessing pipeline changes. The data, the model and the subject all stay the same. We characterize the sensitivity with a Walsh-Hadamard decomposition, introduce Preprocessing Uncertainty as a per-trial diagnostic, and mitigate the instability with Normalized Adaptive PGI.

Read more

Probing Working Memory in LLMs

WMF-AM isolates cumulative state tracking, the ability to maintain and update an intermediate result across K operations with no scratchpad. Across 20 open-weight models from 13 families, the probe predicts downstream agent performance at r = 0.612, and it stays discriminative where fixed benchmarks plateau.

Read more

Frontal-Midline Theta Ramping

Simultaneous EEG and eye-tracking during visual search show frontal-midline theta ramping up before voluntary, self-initiated attention shifts. The ramp indexes attentional preparation and separates self-initiated shifts from externally driven ones.

Read more

News

Education

Teaching

Instructor, Graduate School of Information Sciences, Tohoku University

Teaching Assistant, Graduate School of Information Sciences, Tohoku University (2022 – 2025)

Supervisors: Prof. Kazunori Yamada, Assoc. Prof. Samy Baladram

• Spring 2024: Machine Learning Basics, Data Science Programming Basics, Data Engineering, Data Science Training I & II
• Fall 2023: Data Science Basics
• Spring 2022: Data Science Training I, Data Science Skill Up Exercise

Grants, Fellowships & Awards

Editorial Board

• Editorial Board Member, Interdisciplinary Information Sciences (Tohoku University GSIS), 2026–present

Reviewer

• npj Science of Learning
• Cognitive Neurodynamics
• Journal of NeuroEngineering and Rehabilitation
• Scientific Reports
• Discover Neuroscience
• European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD 2026)
• International Joint Conference on Neural Networks (IJCNN)

Links

• International Liaison Office (ILO)
• Graduate Program in Data Science (GPDS)
• Yamada Laboratory
• TOHOKU AI GROUP, Artificial Intelligence Research Group at Tohoku University
• Graduate School of Information Sciences (GSIS), Tohoku University
• Unprecedented-scale Data Analytics Center (UDAC), Social Integration Research Division
• Collaborative Research Laboratory for Technology Governance (TechGov), Tohoku University × NTT Data Group