Linear Independence of Polynomial Compositions and Identifiability of Deep Neural Networks
研究通过证明多项式组合的线性独立性来解决深度神经网络的可识别性问题，特别针对具有特定激活函数的网络架构。
0 citationsRead paper
Selected work
Representative Papers
Enhanced Filtering Algorithms for the Euclidean Traveling Salesperson Problem and its variants in Constraint Logic Programming
This work addresses the limitation of traditional constraint programming approaches to the Euclidean Traveling Salesman Problem (TSP), which fail to exploit the geometric information inherent in node coordinates, thereby restricting constraint propagation. For the first time, the authors explicitly integrate geometric reasoning into a constraint logic programming framework and devise novel filtering algorithms that fully leverage the properties of Euclidean distances. This integration substantially strengthens constraint propagation and naturally extends to variants such as the Euclidean Generalized TSP. Empirical evaluations demonstrate the method’s superior computational efficiency over existing techniques, particularly in practical applications like route planning and logistics.
0 citationsRead paper
Mahalanobis PatchCore: Covariance-Aware and Streaming-Compatible Industrial Anomaly Detection
This work addresses key challenges in industrial visual anomaly detection—namely, the scarcity of defective samples, neglect of feature correlations, and high memory consumption—by proposing a covariance-aware streaming anomaly detection method. The approach integrates Mahalanobis distance into nearest-neighbor retrieval through regularized covariance modeling and feature whitening. To enable efficient streaming memory bank construction, it introduces a bounded-memory training mechanism supporting incremental principal component analysis and online covariance estimation. The method achieves competitive image-level detection performance while substantially reducing peak memory usage from 5.41 GB to 2.78 GB and attains an average AUC of 0.986 on industrial datasets, significantly outperforming existing solutions.
0 citationsRead paper
Graph Memory Transformer (GMT)
This work proposes the Graph Memory Transformer (GMT), which replaces the feed-forward networks (FFNs) in a standard autoregressive Transformer decoder with an explicit, learnable memory graph while preserving the causal self-attention mechanism. GMT implements interpretable state transitions through a routing-and-displacement scheme based on 128 memory centroids, a 128×128 directed transition matrix, gravity-source routing, token-conditioned target selection, and gated readout, yielding a pure decoder language model without FFNs. With 82.2M parameters, GMT trains stably and achieves validation loss and perplexity slightly behind those of a 103.0M-parameter GPT baseline (3.5995/36.58 vs. 3.2903/26.85), yet demonstrates comparable zero-shot performance, thereby validating the feasibility and interpretability of the graph-based memory mechanism.
0 citationsRead paper
Asymptotic-Preserving Neural Networks for Viscoelastic Parameter Identification in Multiscale Blood Flow Modeling
This study addresses the clinical challenge of directly measuring intravascular pressure by proposing a novel approach that integrates physical principles with deep learning. Specifically, it employs Asymptotic-Preserving Neural Networks (APNN) to simultaneously infer arterial wall viscoelastic parameters and reconstruct the temporal evolution of vascular state variables from Doppler ultrasound-derived cross-sectional area and blood flow velocity data. By embedding asymptotic-preserving properties into the neural network architecture, this method enables end-to-end, physics-consistent learning of parameters in a one-dimensional multiscale viscoelastic blood flow model. Validation on both synthetic and patient-specific datasets demonstrates accurate reconstruction of pressure waveforms, confirming the approach’s effectiveness and robustness in scenarios where direct pressure measurements are unavailable.
0 citationsRead paper
Recent publications
Latest Papers
Linear Independence of Polynomial Compositions and Identifiability of Deep Neural Networks
研究通过证明多项式组合的线性独立性来解决深度神经网络的可识别性问题，特别针对具有特定激活函数的网络架构。
0 citationsRead paper
Enhanced Filtering Algorithms for the Euclidean Traveling Salesperson Problem and its variants in Constraint Logic Programming
This work addresses the limitation of traditional constraint programming approaches to the Euclidean Traveling Salesman Problem (TSP), which fail to exploit the geometric information inherent in node coordinates, thereby restricting constraint propagation. For the first time, the authors explicitly integrate geometric reasoning into a constraint logic programming framework and devise novel filtering algorithms that fully leverage the properties of Euclidean distances. This integration substantially strengthens constraint propagation and naturally extends to variants such as the Euclidean Generalized TSP. Empirical evaluations demonstrate the method’s superior computational efficiency over existing techniques, particularly in practical applications like route planning and logistics.
0 citationsRead paper
Mahalanobis PatchCore: Covariance-Aware and Streaming-Compatible Industrial Anomaly Detection
This work addresses key challenges in industrial visual anomaly detection—namely, the scarcity of defective samples, neglect of feature correlations, and high memory consumption—by proposing a covariance-aware streaming anomaly detection method. The approach integrates Mahalanobis distance into nearest-neighbor retrieval through regularized covariance modeling and feature whitening. To enable efficient streaming memory bank construction, it introduces a bounded-memory training mechanism supporting incremental principal component analysis and online covariance estimation. The method achieves competitive image-level detection performance while substantially reducing peak memory usage from 5.41 GB to 2.78 GB and attains an average AUC of 0.986 on industrial datasets, significantly outperforming existing solutions.
0 citationsRead paper
Graph Memory Transformer (GMT)
This work proposes the Graph Memory Transformer (GMT), which replaces the feed-forward networks (FFNs) in a standard autoregressive Transformer decoder with an explicit, learnable memory graph while preserving the causal self-attention mechanism. GMT implements interpretable state transitions through a routing-and-displacement scheme based on 128 memory centroids, a 128×128 directed transition matrix, gravity-source routing, token-conditioned target selection, and gated readout, yielding a pure decoder language model without FFNs. With 82.2M parameters, GMT trains stably and achieves validation loss and perplexity slightly behind those of a 103.0M-parameter GPT baseline (3.5995/36.58 vs. 3.2903/26.85), yet demonstrates comparable zero-shot performance, thereby validating the feasibility and interpretability of the graph-based memory mechanism.
0 citationsRead paper
Asymptotic-Preserving Neural Networks for Viscoelastic Parameter Identification in Multiscale Blood Flow Modeling
This study addresses the clinical challenge of directly measuring intravascular pressure by proposing a novel approach that integrates physical principles with deep learning. Specifically, it employs Asymptotic-Preserving Neural Networks (APNN) to simultaneously infer arterial wall viscoelastic parameters and reconstruct the temporal evolution of vascular state variables from Doppler ultrasound-derived cross-sectional area and blood flow velocity data. By embedding asymptotic-preserving properties into the neural network architecture, this method enables end-to-end, physics-consistent learning of parameters in a one-dimensional multiscale viscoelastic blood flow model. Validation on both synthetic and patient-specific datasets demonstrates accurate reconstruction of pressure waveforms, confirming the approach’s effectiveness and robustness in scenarios where direct pressure measurements are unavailable.
0 citationsRead paper