Selected work

Representative Papers

EMTSF:Extraordinary Mixture of SOTA Models for Time Series Forecasting

Oct 27, 2025

Challenged by the limited efficacy of Transformers in time-series forecasting (TSF), the insufficient robustness of LLM-based approaches, and the over-dominance of recent observations, this paper proposes the first Transformer-gated Mixture-of-Experts (MoE) framework integrating multiple state-of-the-art paradigms. The framework unifies four heterogeneous models—xLSTM, an enhanced linear model, PatchTST, and minGRU—under a learnable Transformer-based gating network for dynamic expert weighting. It further introduces a recency-prioritized temporal weighting scheme to strengthen local dynamics modeling. Distinct from existing MoE methods, this work achieves cross-architectural complementarity within a single unified architecture, significantly improving both accuracy and robustness. Extensive experiments demonstrate consistent superiority over leading TSF models—including TimeLLM—across multiple standard benchmarks, empirically validating the effectiveness of heterogeneous model collaboration.

0 citationsRead paper

Revolutionizing Wildfire Detection with Convolutional Neural Networks: A VGG16 Model Approach

May 26, 20252025 Northeast Section Conference Proceedings

To address the high false-negative rate and deployment challenges on edge devices in early wildfire detection, this paper proposes a lightweight fine-tuning approach for VGG16 tailored to野外 wildfire imagery. To mitigate the low-resolution and severe class imbalance issues inherent in the D-FIRE dataset, we apply targeted data augmentation—including rotation, scaling, and color jittering—and introduce a confusion-matrix-driven classification threshold optimization strategy to explicitly minimize false negatives. The resulting end-to-end binary classifier achieves real-time inference capability while maintaining high reliability. Evaluated on the D-FIRE benchmark, our model attains 98.2% accuracy and a false-negative rate below 0.5%. This work delivers a practical, edge-deployable solution for high-reliability early wildfire identification under resource-constrained conditions.

0 citationsRead paper

Wildfire Detection Using Vision Transformer with the Wildfire Dataset

May 23, 20252025 Northeast Section Conference Proceedings

To address challenges in wildfire detection—including limited remote sensing coverage, smoke interference, difficulty in detecting small-scale fire targets, and poor model real-time performance—this paper proposes an end-to-end Vision Transformer (ViT)-based approach for wildfire image recognition. We present the first systematic adaptation of ViT architectures to wildfire detection, introducing novel data preprocessing and normalization strategies specifically designed for smoke-affected and low-contrast scenes. The model is trained on a high-resolution, 10.74 GB dataset comprising ‘fire’ and ‘nofire’ images. Evaluated on real-world wildfire data, it achieves 98.2% accuracy and 96.7% recall, with an F1-score 4.1 percentage points higher than ResNet-50. This demonstrates significantly improved robustness in early-stage fire detection and identification of small-scale flames. Moreover, the model supports efficient lightweight deployment on edge devices.

0 citationsRead paper

Neural Attention: A Novel Mechanism for Enhanced Expressive Power in Transformer Models

Feb 24, 2025

Standard Transformer self-attention relies on dot-product similarity, which lacks expressivity for capturing complex, nonlinear token interactions. To address this, we propose Neural Attention—a novel attention mechanism that replaces the fixed dot-product with a learnable feed-forward neural network to compute attention weights. This is the first approach to introduce parameterized nonlinear mapping into attention weight computation while preserving dimensional compatibility and mathematical differentiability, thereby substantially enhancing representational capacity. The mechanism is inherently cross-modal: it achieves consistent gains across NLP and CV benchmarks—reducing perplexity by over 5% on WikiText-103 and significantly improving classification accuracy on CIFAR-10/100. Through rigorous computational complexity analysis and architectural optimization, we ensure both enhanced modeling capability and practical deployability.

0 citationsRead paper

Recent publications

Latest Papers

EMTSF:Extraordinary Mixture of SOTA Models for Time Series Forecasting

Oct 27, 2025

Challenged by the limited efficacy of Transformers in time-series forecasting (TSF), the insufficient robustness of LLM-based approaches, and the over-dominance of recent observations, this paper proposes the first Transformer-gated Mixture-of-Experts (MoE) framework integrating multiple state-of-the-art paradigms. The framework unifies four heterogeneous models—xLSTM, an enhanced linear model, PatchTST, and minGRU—under a learnable Transformer-based gating network for dynamic expert weighting. It further introduces a recency-prioritized temporal weighting scheme to strengthen local dynamics modeling. Distinct from existing MoE methods, this work achieves cross-architectural complementarity within a single unified architecture, significantly improving both accuracy and robustness. Extensive experiments demonstrate consistent superiority over leading TSF models—including TimeLLM—across multiple standard benchmarks, empirically validating the effectiveness of heterogeneous model collaboration.

0 citationsRead paper

Revolutionizing Wildfire Detection with Convolutional Neural Networks: A VGG16 Model Approach

May 26, 20252025 Northeast Section Conference Proceedings

To address the high false-negative rate and deployment challenges on edge devices in early wildfire detection, this paper proposes a lightweight fine-tuning approach for VGG16 tailored to野外 wildfire imagery. To mitigate the low-resolution and severe class imbalance issues inherent in the D-FIRE dataset, we apply targeted data augmentation—including rotation, scaling, and color jittering—and introduce a confusion-matrix-driven classification threshold optimization strategy to explicitly minimize false negatives. The resulting end-to-end binary classifier achieves real-time inference capability while maintaining high reliability. Evaluated on the D-FIRE benchmark, our model attains 98.2% accuracy and a false-negative rate below 0.5%. This work delivers a practical, edge-deployable solution for high-reliability early wildfire identification under resource-constrained conditions.

0 citationsRead paper

Wildfire Detection Using Vision Transformer with the Wildfire Dataset

May 23, 20252025 Northeast Section Conference Proceedings

To address challenges in wildfire detection—including limited remote sensing coverage, smoke interference, difficulty in detecting small-scale fire targets, and poor model real-time performance—this paper proposes an end-to-end Vision Transformer (ViT)-based approach for wildfire image recognition. We present the first systematic adaptation of ViT architectures to wildfire detection, introducing novel data preprocessing and normalization strategies specifically designed for smoke-affected and low-contrast scenes. The model is trained on a high-resolution, 10.74 GB dataset comprising ‘fire’ and ‘nofire’ images. Evaluated on real-world wildfire data, it achieves 98.2% accuracy and 96.7% recall, with an F1-score 4.1 percentage points higher than ResNet-50. This demonstrates significantly improved robustness in early-stage fire detection and identification of small-scale flames. Moreover, the model supports efficient lightweight deployment on edge devices.

0 citationsRead paper

Neural Attention: A Novel Mechanism for Enhanced Expressive Power in Transformer Models

Feb 24, 2025

Standard Transformer self-attention relies on dot-product similarity, which lacks expressivity for capturing complex, nonlinear token interactions. To address this, we propose Neural Attention—a novel attention mechanism that replaces the fixed dot-product with a learnable feed-forward neural network to compute attention weights. This is the first approach to introduce parameterized nonlinear mapping into attention weight computation while preserving dimensional compatibility and mathematical differentiability, thereby substantially enhancing representational capacity. The mechanism is inherently cross-modal: it achieves consistent gains across NLP and CV benchmarks—reducing perplexity by over 5% on WikiText-103 and significantly improving classification accuracy on CIFAR-10/100. Through rigorous computational complexity analysis and architectural optimization, we ensure both enhanced modeling capability and practical deployability.

0 citationsRead paper