Skip to main content
Usami Laboratory

Department of Computer Science, College of Engineering, Chubu University

Usami Laboratory

Understanding light and the principles of observation and inference

We study how light interacts with objects and environments, and what information images can reveal. Mathematical analysis and reproducible experiments examine when measurement relationships and estimates hold, informing the validation of AI estimation, generation, and evaluation.

Selected Research

Selected research

The questions, methods, and evaluation results behind selected studies.

LLM Evaluation & Reliability

Testing LLM judges as measuring instruments

We test LLM judges as measuring instruments: whether verdicts can be read as specified, whether identical answers receive non-tie judgments, and how accurately a target is selected under declared criteria.

LLM Judges Have Dark Current · arXiv preprint · 2026

Joint research with Mitsubishi Heavy Industries, Research & Innovation Center

Read research overview

Measurement stages

Output validity
Can the declared parser read the verdict?
Dark current
Does the judge prefer one exact duplicate?
Target-selection utility
Does it select the specified target?
All research explainers

Recent preprints

Current arXiv versions of studies in machine learning and polarization estimation. These preprints have not been peer reviewed.

Preprint · Not peer reviewed

Exact Fusion of Softmax Attention Heads Can Require Exponentially Many Atoms

The paper constructs independently normalized attention heads whose exact fusion into one head requires exponentially more key-value entries. The preprint and reproduction materials are available.

Hiroyasu Usami, Keisuke Hara, Ayato Tsuboi, Naohiko Matsuda

Twelve separately normalized two-atom heads, with 24 head–atom terms, compared with 4,096 atoms in one exact single-head readout.
In this constructed example, 12 two-atom heads use 24 head–atom terms; an exactly equivalent single-head representation requires at least 4,096 atoms.
Preprint · Not peer reviewed

Conditional stopping guarantees for a folded parametrization of reduced Stokes vectors

Derives an objective-gap bound from derivative and curvature tests for strongly convex physical objectives on the bounded-intensity inner sheet of a folded Stokes parametrization, including partial linear polarization.

Hiroyasu Usami, Ayato Tsuboi, Keisuke Hara

The fold in a Stokes parametrization and the additional conditions for an inner-sheet stopping certificate.
An English-language animation explaining the Stokes fold. Stopping guarantees require the convexity, domain, and error conditions specified in the paper.
Preprint · Not peer reviewed

Invertible Graph Aggregation Can Increase ReLU Readout Width

Studies how invertible graph aggregation can increase the width needed by a shared nodewise ReLU classifier with one hidden layer, using an explicit tree construction, perturbation and noise bounds, controlled fits, and public-graph comparisons.

Hiroyasu Usami, Masanori Tsukamoto, Kohei Hashiguchi

Constructed 24-node tree and scalar feature orders before and after averaging. Raw features separate at one threshold; aggregated leaf labels alternate.
In this constructed tree (N = 12), one invertible averaging step raises the minimum width of a shared one-hidden-layer ReLU classifier from 1 to 12 for zero classification error.

Laboratory record

Outputs are site-listed records from 2022 onward; awards and patents are counted from this site's activity records. Site data checked: 2026-11-04. These counts differ in scope from the PI's university-database record.

Hiroyasu Usami — profile

News

Research presentations, student activities, awards, and academic service.

Research and laboratory activities

Academic service

Academic Service

Joined the Editorial Board of Scientific Reports

Hiroyasu Usami joined the Editorial Board of Scientific Reports, a Nature Portfolio journal, in September 2026, with a focus on computer vision, machine learning, and medical image analysis.

View all activity

People

Faculty, graduate students, and undergraduates work together on image understanding and AI.

1

Faculty

25

Current Students

21

Alumni

Selected Career Paths

Companies are listed in Japanese alphabetical order.

  • AISIN Digital Engineering Co., Ltd.
  • Aiphone Co., Ltd.
  • ARGO GRAPHICS Inc.
  • Enshu Limited
  • Gothic Co., Ltd.
  • Common Products Co., Ltd.
  • C-Cube Co., Ltd.
  • J:COM Co., Ltd.
  • JBCC Corporation
  • Sumitomo Riko Company Limited
  • TechPresso Inc.
  • DIC Solution Engineering Co., Ltd.
  • Nodajuku Co., Ltd.

Graduate study: Chubu University Graduate School

For Prospective Students

Research topics, joining procedures, and information on laboratory visits and consultations.

For Research Collaboration

Contact us about collaborative research in image analysis, machine learning, and LLM evaluation.