Selected work

Representative Papers

Decorrelation Speeds Up Vision Transformers

Oct 16, 2025

Vision Transformers (ViTs) pretrained via Masked Autoencoders (MAEs) achieve strong performance under low-label regimes, yet their high computational cost hinders industrial deployment. To address this, we propose **Selective Decorrelation Backpropagation (DecorrBP)**—a lightweight optimization technique that imposes layer-wise gradient covariance constraints exclusively within the MAE encoder, enhancing gradient propagation efficiency and convergence speed while preserving training stability. Evaluated on ImageNet-1K, DecorrBP reduces pretraining time by 21.1% and carbon emissions by 21.4%. On downstream ADE20K semantic segmentation, it improves mIoU by 1.1 points; consistent gains are also observed on industrial datasets. Crucially, DecorrBP is the first method to integrate gradient decorrelation into the MAE training framework without modifying model architecture or loss functions—enabling efficient, low-carbon, and high-performance ViT pretraining.

0 citationsRead paper

Interactive Semantic Segmentation for Phosphene Vision Neuroprosthetics

Sep 24, 2025

Low phosphene resolution and semantic ambiguity in visual neuroprostheses hinder object recognition. To address this, we propose a user-centered, gaze-guided semantic segmentation framework. Methodologically, we introduce the Segment Anything Model (SAM) into phosphene vision simulation for the first time, integrating real-time eye-tracking with edge detection to enable interactive, goal-directed segmentation under simplified visual conditions. Our contributions are twofold: (1) a novel gaze-guided dynamic focusing mechanism that enhances alignment between user intent and segmentation output; and (2) empirical validation of SAM’s robustness in segmenting irregular and shape-specific objects from low-resolution, high-noise phosphene images. Experiments demonstrate a 23.6% average improvement in recognition accuracy over conventional edge-based methods, significantly enhancing semantic interpretability and interaction efficiency in complex scenes. This work establishes a new paradigm for vision augmentation in neural prosthetics.

0 citationsRead paper

eegFloss: A Python package for refining sleep EEG recordings using machine learning models

Jul 08, 2025

Sleep EEG analysis is highly susceptible to both endogenous device-related and exogenous environmental noise, leading to erroneous automatic sleep staging and inaccurate scoring. To address this, we propose an end-to-end, open-source quality control framework. First, we introduce *eegUsability*, a deep learning model trained on multi-subject, multi-night manually annotated data, achieving high recall (94%) for usable signal detection and strong cross-subject generalizability (F1 = 0.85, Cohen’s κ = 0.78). Second, we develop *eegMobility*, a model enabling fully automated bed-time detection. Together, these components enable integrated artifact filtering, sleep staging, and statistical analysis while maintaining compatibility across diverse EEG acquisition devices. The framework significantly enhances data reliability and automatic analysis accuracy in large-scale sleep studies.

0 citationsRead paper

RocketStack: A level-aware deep recursive ensemble learning framework with exploratory feature fusion and model pruning dynamics

Jun 20, 2025

Deep stacking ensembles are hindered by excessive model complexity, feature redundancy, and high computational overhead, limiting their ability to surpass shallow-layer performance ceilings. To address this, RocketStack introduces a hierarchy-aware deep stacking framework enabling up to ten levels of recursive stacking. It incorporates three novel components: (i) hierarchical adaptive pruning, (ii) lightweight Gaussian perturbation regularization, and (iii) a periodic collaborative compression mechanism integrating attention, spectral feature extraction (SFE), and autoencoders. This establishes a “prune–compress–propagate” tri-level optimization pipeline. The method effectively alleviates early performance saturation while enhancing generalization and inference efficiency: binary classification accuracy improves to 97.08% (+5.14%), and multiclass accuracy reaches 98.60% (+6.11%). Feature dimensionality is reduced by 74%, and inference latency decreases by 56.1%.

0 citationsRead paper

Recent publications

Latest Papers

Decorrelation Speeds Up Vision Transformers

Oct 16, 2025

Vision Transformers (ViTs) pretrained via Masked Autoencoders (MAEs) achieve strong performance under low-label regimes, yet their high computational cost hinders industrial deployment. To address this, we propose **Selective Decorrelation Backpropagation (DecorrBP)**—a lightweight optimization technique that imposes layer-wise gradient covariance constraints exclusively within the MAE encoder, enhancing gradient propagation efficiency and convergence speed while preserving training stability. Evaluated on ImageNet-1K, DecorrBP reduces pretraining time by 21.1% and carbon emissions by 21.4%. On downstream ADE20K semantic segmentation, it improves mIoU by 1.1 points; consistent gains are also observed on industrial datasets. Crucially, DecorrBP is the first method to integrate gradient decorrelation into the MAE training framework without modifying model architecture or loss functions—enabling efficient, low-carbon, and high-performance ViT pretraining.

0 citationsRead paper

Interactive Semantic Segmentation for Phosphene Vision Neuroprosthetics

Sep 24, 2025

Low phosphene resolution and semantic ambiguity in visual neuroprostheses hinder object recognition. To address this, we propose a user-centered, gaze-guided semantic segmentation framework. Methodologically, we introduce the Segment Anything Model (SAM) into phosphene vision simulation for the first time, integrating real-time eye-tracking with edge detection to enable interactive, goal-directed segmentation under simplified visual conditions. Our contributions are twofold: (1) a novel gaze-guided dynamic focusing mechanism that enhances alignment between user intent and segmentation output; and (2) empirical validation of SAM’s robustness in segmenting irregular and shape-specific objects from low-resolution, high-noise phosphene images. Experiments demonstrate a 23.6% average improvement in recognition accuracy over conventional edge-based methods, significantly enhancing semantic interpretability and interaction efficiency in complex scenes. This work establishes a new paradigm for vision augmentation in neural prosthetics.

0 citationsRead paper

eegFloss: A Python package for refining sleep EEG recordings using machine learning models

Jul 08, 2025

Sleep EEG analysis is highly susceptible to both endogenous device-related and exogenous environmental noise, leading to erroneous automatic sleep staging and inaccurate scoring. To address this, we propose an end-to-end, open-source quality control framework. First, we introduce *eegUsability*, a deep learning model trained on multi-subject, multi-night manually annotated data, achieving high recall (94%) for usable signal detection and strong cross-subject generalizability (F1 = 0.85, Cohen’s κ = 0.78). Second, we develop *eegMobility*, a model enabling fully automated bed-time detection. Together, these components enable integrated artifact filtering, sleep staging, and statistical analysis while maintaining compatibility across diverse EEG acquisition devices. The framework significantly enhances data reliability and automatic analysis accuracy in large-scale sleep studies.

0 citationsRead paper

RocketStack: A level-aware deep recursive ensemble learning framework with exploratory feature fusion and model pruning dynamics

Jun 20, 2025

Deep stacking ensembles are hindered by excessive model complexity, feature redundancy, and high computational overhead, limiting their ability to surpass shallow-layer performance ceilings. To address this, RocketStack introduces a hierarchy-aware deep stacking framework enabling up to ten levels of recursive stacking. It incorporates three novel components: (i) hierarchical adaptive pruning, (ii) lightweight Gaussian perturbation regularization, and (iii) a periodic collaborative compression mechanism integrating attention, spectral feature extraction (SFE), and autoencoders. This establishes a “prune–compress–propagate” tri-level optimization pipeline. The method effectively alleviates early performance saturation while enhancing generalization and inference efficiency: binary classification accuracy improves to 97.08% (+5.14%), and multiclass accuracy reaches 98.60% (+6.11%). Feature dimensionality is reduced by 74%, and inference latency decreases by 56.1%.

0 citationsRead paper