Selected work

Representative Papers

Smallest suffixient set maintenance in near-real-time

Apr 30, 2026

This work addresses the problem of dynamically maintaining a minimal suffix-rich set for a string under near-real-time constraints to efficiently quantify its repetitiveness. Focusing on online scenarios where characters arrive one by one—either left-to-right or right-to-left—it presents the first algorithm achieving polyloglog worst-case time per character update. The approach leverages Weiner’s suffix tree and its fundamental algorithmic primitives to establish a core maintenance mechanism, thereby enabling, for the first time, efficient dynamic maintenance of minimal suffix-rich sets under bidirectional streaming input. This breakthrough substantially extends the applicability of string repetitiveness measures to dynamic environments.

0 citationsRead paper

Data-Efficient Semantic Segmentation of 3D Point Clouds via Open-Vocabulary Image Segmentation-based Pseudo-Labeling

Apr 13, 2026

This work addresses the triple challenge of scarce training scenes, limited point-level annotations, and absent corresponding 2D images in 3D point cloud semantic segmentation by proposing the PLOVIS framework. PLOVIS is the first method to jointly tackle these issues without relying on real 2D images; instead, it leverages open-vocabulary image segmentation models to generate high-quality 2D pseudo-labels from 3D point clouds. To enhance training efficiency, the framework incorporates a two-stage pseudo-label filtering mechanism and a class-balanced memory bank. Evaluated on four standard benchmarks—ScanNet, S3DIS, Toronto3D, and Semantic3D—PLOVIS achieves state-of-the-art performance using only dozens of scenes with fewer than 100 annotated points per scene, significantly outperforming existing approaches.

0 citationsRead paper

Near-real-time Solutions for Online String Problems

Feb 16, 2026

This work presents the first online algorithms achieving worst-case poly-log-logarithmic time per character for several classical problems in online string processing, including maintaining the longest repeated suffix array, (reverse) LZ77 factorization, and minimal unique substrings. Building upon the Breslauer–Italiano online suffix tree construction and leveraging irreducible representations of string data structures, the proposed approach operates within linear space while significantly improving upon prior methods that only offered amortized bounds or higher logarithmic complexities. Furthermore, the study establishes an efficient transformation between the longest previous factor array and the longest repeated suffix array, enabling sublinear-sized updates to these arrays.

0 citationsRead paper

Motional representation; the ability to predict odor characters using molecular vibrations

Sep 17, 2025

Predicting odor perception from molecular structure remains a fundamental challenge. This work introduces CNN_vib, a novel convolutional neural network regression model that systematically investigates molecular vibrational spectra—as opposed to conventional static structural representations—as a paradigm for odor prediction. We construct a parameterized vibrational spectral representation and benchmark it against molecular fingerprints and logistic regression across multiple odor descriptors (e.g., “sweet,” “pungent,” “woody”). Results demonstrate that vibrational spectra achieve predictive performance comparable to or exceeding that of molecular fingerprints. Crucially, we show that molecular dynamic features—particularly low-frequency vibrational modes—encode essential olfactory information sufficient for odor prediction in isolation. CNN_vib significantly enhances the modeling capacity for vibrational spectral data. This study breaks the long-standing reliance on static molecular structures in computational olfaction, establishing vibrational spectroscopy as a theoretically grounded and technically viable foundation for odor prediction.

0 citationsRead paper

Token Adaptation via Side Graph Convolution for Temporally and Spatially Efficient Fine-tuning of 3D Point Cloud Transformers

Feb 19, 2025

To address the high computational overhead and weak adaptation capability in parameter-efficient fine-tuning (PEFT) of 3D point cloud Transformers, this paper proposes STAG—a token dynamic adaptation method leveraging parallel side graph convolutional networks. Its core innovations include: (i) the first point-cloud-specific side graph convolutional token adaptation mechanism; (ii) gradient-thinned connections; (iii) cross-layer parameter sharing; and (iv) lightweight graph convolutions. For systematic evaluation, we introduce PCC13—the first PEFT-dedicated point cloud benchmark, comprising 13 diverse datasets. Experiments demonstrate that STAG achieves classification accuracy on par with full-parameter fine-tuning while requiring only 0.43M trainable parameters. It reduces GPU memory consumption by 62% and training time by 57%. Moreover, STAG consistently outperforms existing PEFT methods across multiple tasks in PCC13, establishing new state-of-the-art performance for efficient point cloud Transformer adaptation.

0 citationsRead paper

Recent publications

Latest Papers

Smallest suffixient set maintenance in near-real-time

Apr 30, 2026

This work addresses the problem of dynamically maintaining a minimal suffix-rich set for a string under near-real-time constraints to efficiently quantify its repetitiveness. Focusing on online scenarios where characters arrive one by one—either left-to-right or right-to-left—it presents the first algorithm achieving polyloglog worst-case time per character update. The approach leverages Weiner’s suffix tree and its fundamental algorithmic primitives to establish a core maintenance mechanism, thereby enabling, for the first time, efficient dynamic maintenance of minimal suffix-rich sets under bidirectional streaming input. This breakthrough substantially extends the applicability of string repetitiveness measures to dynamic environments.

0 citationsRead paper

Data-Efficient Semantic Segmentation of 3D Point Clouds via Open-Vocabulary Image Segmentation-based Pseudo-Labeling

Apr 13, 2026

This work addresses the triple challenge of scarce training scenes, limited point-level annotations, and absent corresponding 2D images in 3D point cloud semantic segmentation by proposing the PLOVIS framework. PLOVIS is the first method to jointly tackle these issues without relying on real 2D images; instead, it leverages open-vocabulary image segmentation models to generate high-quality 2D pseudo-labels from 3D point clouds. To enhance training efficiency, the framework incorporates a two-stage pseudo-label filtering mechanism and a class-balanced memory bank. Evaluated on four standard benchmarks—ScanNet, S3DIS, Toronto3D, and Semantic3D—PLOVIS achieves state-of-the-art performance using only dozens of scenes with fewer than 100 annotated points per scene, significantly outperforming existing approaches.

0 citationsRead paper

Near-real-time Solutions for Online String Problems

Feb 16, 2026

This work presents the first online algorithms achieving worst-case poly-log-logarithmic time per character for several classical problems in online string processing, including maintaining the longest repeated suffix array, (reverse) LZ77 factorization, and minimal unique substrings. Building upon the Breslauer–Italiano online suffix tree construction and leveraging irreducible representations of string data structures, the proposed approach operates within linear space while significantly improving upon prior methods that only offered amortized bounds or higher logarithmic complexities. Furthermore, the study establishes an efficient transformation between the longest previous factor array and the longest repeated suffix array, enabling sublinear-sized updates to these arrays.

0 citationsRead paper

Motional representation; the ability to predict odor characters using molecular vibrations

Sep 17, 2025

Predicting odor perception from molecular structure remains a fundamental challenge. This work introduces CNN_vib, a novel convolutional neural network regression model that systematically investigates molecular vibrational spectra—as opposed to conventional static structural representations—as a paradigm for odor prediction. We construct a parameterized vibrational spectral representation and benchmark it against molecular fingerprints and logistic regression across multiple odor descriptors (e.g., “sweet,” “pungent,” “woody”). Results demonstrate that vibrational spectra achieve predictive performance comparable to or exceeding that of molecular fingerprints. Crucially, we show that molecular dynamic features—particularly low-frequency vibrational modes—encode essential olfactory information sufficient for odor prediction in isolation. CNN_vib significantly enhances the modeling capacity for vibrational spectral data. This study breaks the long-standing reliance on static molecular structures in computational olfaction, establishing vibrational spectroscopy as a theoretically grounded and technically viable foundation for odor prediction.

0 citationsRead paper

Token Adaptation via Side Graph Convolution for Temporally and Spatially Efficient Fine-tuning of 3D Point Cloud Transformers

Feb 19, 2025

To address the high computational overhead and weak adaptation capability in parameter-efficient fine-tuning (PEFT) of 3D point cloud Transformers, this paper proposes STAG—a token dynamic adaptation method leveraging parallel side graph convolutional networks. Its core innovations include: (i) the first point-cloud-specific side graph convolutional token adaptation mechanism; (ii) gradient-thinned connections; (iii) cross-layer parameter sharing; and (iv) lightweight graph convolutions. For systematic evaluation, we introduce PCC13—the first PEFT-dedicated point cloud benchmark, comprising 13 diverse datasets. Experiments demonstrate that STAG achieves classification accuracy on par with full-parameter fine-tuning while requiring only 0.43M trainable parameters. It reduces GPU memory consumption by 62% and training time by 57%. Moreover, STAG consistently outperforms existing PEFT methods across multiple tasks in PCC13, establishing new state-of-the-art performance for efficient point cloud Transformer adaptation.

0 citationsRead paper