Deep learning excels at capturing complex data representations, yet quantifying the discriminative quality of these representations remains challenging. While unsupervised metrics often assess pairwise sample similarity, classification tasks fundamentally require class-level discrimination. To bridge this gap, we propose a novel loss function that evaluates representation discriminability via the Euclidean distance between the learned similarity matrix and the true class adjacency matrix. We identify random consistency—an inherent bias in Euclidean distance metrics—as a key obstacle to reliable evaluation, affecting both fairness and discrimination. To address this, we derive the expected Euclidean distance under uniformly distributed label permutations and introduce its closed-form solution, the Pure Square Euclidean Distance (PSED), which provably eliminates random consistency. Theoretically, we demonstrate that PSED satisfies heterogeneity and unbiasedness guarantees, and establish its generalization bound via the exponential Orlicz norm, confirming its statistical learnability. Empirically, our method surpasses conventional loss functions across multiple benchmarks, achieving significant improvements in accuracy, $F_1$ score, and class-structure differentiation. (Code is published in https://github.com/FeijiangLi/ICML2025-PSED)

Virtual Power Plants (VPPs) have gained prominence as an effective approach to enhance renewable energy utilization and optimize power market mechanisms in recent years. To address the inherent uncertainties of renewable energy generation and improve the economic performance of VPPs in electricity markets, this paper proposes a two-stage economic dispatch model that explicitly incorporates renewable energy output forecasting errors. The day-ahead stage employs a data-driven chance-constrained approach to handle renewable uncertainties, ensuring reliable day-ahead scheduling. The real-time stage compensates for deviations between actual and scheduled generation outputs. The proposed coordinated two-stage optimization achieves optimal day-ahead bidding strategies that enhance VPP profitability and reduce real-time power deviations, as validated through case studies.

Quantifying the informativeness of the similarity matrix has been successfully applied in kernel width selection, dimension size selection, and so on. Traditionally, the informational content is defined by calculating the distance between a matrix and non informative matrices. However, this method has limitations when dealing with adjacency matrices with the same marginal distribution but different structures, as it assigns the same distance values to these matrices. To address this issue, we introduce a new distance metric based on the adjacency matrix generated by label vectors, which accurately reflects the correct similarity structure. Through experimental analysis, we have demonstrated that the proposed distance metric can effectively distinguish matrices with different adjacency structures. This feature is crucial for capturing data diversity as it provides a stable measure of sample similarity for classification tasks. In graph neural networks (GNNs), reconstruction loss is crucial for model performance. To verify the effectiveness of the proposed metric, we apply it as a loss function in the training process of the GNN model. Extensive experimental results have shown that the proposed loss function can achieve higher accuracy compared to traditional loss functions.

Masking is widely considered as an effective countermeasure against side-channel analysis (SCA) due to its provable security and efficiency. However, recent works have demonstrated that the deep learning-based SCA (DL-SCA) can effectively break the cryptographic implementations protected by the first-order Boolean maskings. Still, it is open whether higher-order masking can resist DL-SCA. In this work, we demonstrate that deep learning methods can also effectively exploit the inherent leakage of higher-order Boolean masking to compromise its security. Furthermore, we employed neural weight visualization techniques to demonstrate the neural network’s capability to extract high-level features. We assess the efficiency of this novel profiling attack in both simulated and real-world scenarios. In particular, our results show that DL-SCA can effectively break the higher-order Boolean masking schemes up to the sixth and the third order in simulated and real-world cases, respectively. Furthermore, we find that using plaintext-related leakage can significantly improve the effectiveness of side-channel attacks.

Abstract Traditional dehazing algorithms have long been limited in their ability to remove fog effectively, especially in preserving details. This study proposes an improved dark channel prior dehazing algorithm designed to overcome the limitations of traditional algorithms by refining the transmission map and estimating atmospheric light intensity. The algorithm fully exploits the anti‐haze characteristics of the Gabor filter to extract multi‐directional texture features. By fusing these features to form a guidance map for guided filtering, it effectively reduces the blurring caused by guided filtering on the image edges, thereby producing a more accurate transmission map. Simultaneously, an enhanced atmospheric light successfully reduces interference from white objects in the image. The experiment phase utilized the publicly available RESIDE dataset for validation. The algorithm achieved a Peak Signal‐to‐Noise Ratio (PSNR) of 22.4 and a Structural Similarity Index (SSIM) of 0.88. These metrics indicate the algorithm's superior dehazing capabilities.

By separating huge dimensional matrix-matrix multiplication at a single computing node into parallel small matrix multiplications (with appropriate encoding) at parallel worker nodes, coded distributed computing (CDC) tackles the straggler problem and hence speeds up the computation significantly. Existing CDC encoding schemes are based on linear combination (LC), which have two drawbacks: First, heavy computational burden is introduced to both encoding and decoding phases. Second, large numerical error occurs in the decoding phase. To relieve these two effects, a fresh new 2D-SAZD-CDC framework that non-trivially generalizes 1D-SAZD-CDC is proposed, where D is short for dimension, the operation for encoding and decoding is implemented by shift-and-add (SA) and zigzag decoding (ZD) that replaces LC and matrix inversion, respectively. The non-trivial generalization lies in joint design of the operations in 2D are needed in both the encoding and the decoding phases, so as to ensure possesion of combination property (CP) and ZD from 2D viewpoint. More specifically, 2D-SA encoding is designed, 2D-ZD decoding (alternates intermittently between 2D) is proposed, and a proof for satisfying CP and ZD from 2D viewpoint is also given. Numerical studies show that 2D-SAZD-CDC significantly improves the numerical stability and computational load performance over existing LC based schemes.

Energy Management Strategy (EMS) is a crucial part of More-Electric Aircraft (MEA) and aims at improving the efficiency of whole hybrid energy system. In this paper, fuzzy neural network trained by dynamic programming (FNDP) is proposed to solve the problem that dynamic programming (DP) cannot be used online. FNDP can obtain the optimal distribution scheme using DP and extract the rule from the data through fuzzy neural network (FNN). Compared to fuzzy control strategy based on power follow (PFF) in two different load profiles, it can be found that FNDP can attain a satisfied rule without manual setting and have a better performance than PFF.

In this paper, the hybrid power system of fuel cell Unmanned Aerial Vehicle (UAV) is established on the platform of Matlab/Simulink whose system architecture is fuel cell with lithium battery, and the overall scheme of the system is designed. In addition, the required power curves under two typical working conditions are obtained by establishing the aircraft dynamics model. An optimization energy management strategy based on dynamic programming is designed to achieve the least hydrogen consumption. For comparison, another two energy management strategies are designed under the same working condition. Finally, it is proved that dynamic programming has optimal fuel economy than the other two strategies. Moreover, each strategy has its own characteristics, which provides suggestions for selection of energy management strategies under different objectives.

In order to mitigate the environmental pollution problem, the fuel cell electric vehicle (FCEV), as one kind of renewable energy industry, is researched and developed. The performance of FCEV is deeply relied on the energy management strategy (EMS), an inappropriate EMS may cause the bad performance of FCEV. Strategies designed under specific work condition have poor adaptability to complex work conditions. What’s more, these strategies give little concern on the volatility of fuel cell’s output power, which will cause a loss of fuel cell’s life. Aiming at decreasing the hydrogen consumption of FCEV and the volatility of the output power of fuel cell, an optimized fuzzy logic control energy management strategy based on work condition recognition is proposed. In order to achieve a better performance, this strategy can recognize instant driving condition, and pick the corresponding parameters of fuzzy logic control system from the preset database of the parameters. The database is obtained by offline optimization with genetic algorithm. A simulation result based on the MATLAB/Simulink platform is obtained to demonstrate that this strategy has a better performance than a conventional fuzzy logic controller with fixed parameters under mixed work conditions.