Selected work

Representative Papers

Glioma C6: A Novel Dataset for Training and Benchmarking Cell Segmentation

Nov 10, 2025

High-quality annotated data for glioma C6 cell instance segmentation is scarce, hindering robust model development and evaluation. Method: We introduce C6Seg—the first open-source, biologist-curated instance segmentation dataset for C6 cells—comprising 75 phase-contrast microscopy images with over 12,000 pixel-accurate cell masks. C6Seg uniquely incorporates morphological classification labels and subcellular annotations (soma vs. pseudopodia), and spans controlled conditions and multi-condition imaging environments to enhance generalizability. Contribution/Results: Using C6Seg, we systematically benchmark state-of-the-art models (e.g., Mask R-CNN, U-Net), revealing performance bottlenecks in highly clustered and small-object scenarios. Transfer learning on C6Seg yields an 8.2% mAP improvement, validating its utility for model refinement and benchmark establishment. C6Seg thus provides a reproducible, high-fidelity resource for quantitative analysis of brain tumor cells.

0 citationsRead paper

Neural network task specialization via domain constraining

Apr 28, 2025

This paper addresses the suboptimal performance of general-purpose neural networks on specific data subspaces (e.g., image classification and object detection) by proposing a task-oriented domain-constrained specialization method that requires no additional data and preserves the original training pipeline. Methodologically, it introduces a novel two-stage “expert extraction–pre-tuning” mechanism; identifies semantic-consistent subspace constraints as critical for performance gains; and integrates feature-space evolution analysis, class-label space pruning, and an enhanced fine-tuning strategy to build a dynamic, configurable analytical system. Experimental results demonstrate that constraining only the input domain—without architectural or training modifications—significantly improves accuracy across mainstream models. The approach validates the effectiveness and generalizability of domain-constraint-driven model specialization, offering a principled pathway toward efficient, task-specific adaptation of foundation models.

0 citationsRead paper

Hypernym Bias: Unraveling Deep Classifier Training Dynamics through the Lens of Class Hierarchy

Feb 17, 2025

This work investigates how deep classifiers dynamically model semantic hierarchies among classes during training. To address this, we propose the first training dynamics analysis framework grounded in class hierarchy evolution. Our method integrates hierarchy-aware visualization, cross-layer representation tracking, label clustering metrics, and quantification of neural collapse. It reveals that feature manifolds progressively align with hypernym–hyponym semantic structures across network layers: early layers prioritize separation of coarse-grained (hypernym) categories, while deeper layers refine fine-grained (hyponym) distinctions; moreover, neural collapse occurs earlier in the hypernym label space. Experiments demonstrate that deep networks inherently perform hierarchical learning aligned with the intrinsic data hierarchy, yielding feature representations highly consistent with ground-truth class taxonomy. This significantly enhances the interpretability of training dynamics on standard benchmarks such as ImageNet.

0 citationsRead paper

Recent publications

Latest Papers

Glioma C6: A Novel Dataset for Training and Benchmarking Cell Segmentation

Nov 10, 2025

High-quality annotated data for glioma C6 cell instance segmentation is scarce, hindering robust model development and evaluation. Method: We introduce C6Seg—the first open-source, biologist-curated instance segmentation dataset for C6 cells—comprising 75 phase-contrast microscopy images with over 12,000 pixel-accurate cell masks. C6Seg uniquely incorporates morphological classification labels and subcellular annotations (soma vs. pseudopodia), and spans controlled conditions and multi-condition imaging environments to enhance generalizability. Contribution/Results: Using C6Seg, we systematically benchmark state-of-the-art models (e.g., Mask R-CNN, U-Net), revealing performance bottlenecks in highly clustered and small-object scenarios. Transfer learning on C6Seg yields an 8.2% mAP improvement, validating its utility for model refinement and benchmark establishment. C6Seg thus provides a reproducible, high-fidelity resource for quantitative analysis of brain tumor cells.

0 citationsRead paper

Neural network task specialization via domain constraining

Apr 28, 2025

This paper addresses the suboptimal performance of general-purpose neural networks on specific data subspaces (e.g., image classification and object detection) by proposing a task-oriented domain-constrained specialization method that requires no additional data and preserves the original training pipeline. Methodologically, it introduces a novel two-stage “expert extraction–pre-tuning” mechanism; identifies semantic-consistent subspace constraints as critical for performance gains; and integrates feature-space evolution analysis, class-label space pruning, and an enhanced fine-tuning strategy to build a dynamic, configurable analytical system. Experimental results demonstrate that constraining only the input domain—without architectural or training modifications—significantly improves accuracy across mainstream models. The approach validates the effectiveness and generalizability of domain-constraint-driven model specialization, offering a principled pathway toward efficient, task-specific adaptation of foundation models.

0 citationsRead paper

Hypernym Bias: Unraveling Deep Classifier Training Dynamics through the Lens of Class Hierarchy

Feb 17, 2025

This work investigates how deep classifiers dynamically model semantic hierarchies among classes during training. To address this, we propose the first training dynamics analysis framework grounded in class hierarchy evolution. Our method integrates hierarchy-aware visualization, cross-layer representation tracking, label clustering metrics, and quantification of neural collapse. It reveals that feature manifolds progressively align with hypernym–hyponym semantic structures across network layers: early layers prioritize separation of coarse-grained (hypernym) categories, while deeper layers refine fine-grained (hyponym) distinctions; moreover, neural collapse occurs earlier in the hypernym label space. Experiments demonstrate that deep networks inherently perform hierarchical learning aligned with the intrinsic data hierarchy, yielding feature representations highly consistent with ground-truth class taxonomy. This significantly enhances the interpretability of training dynamics on standard benchmarks such as ImageNet.

0 citationsRead paper