• We propose a novel Actor-Critic model with PPO for solving the TSP. • Our adaptive scheduling method improves learning efficiency and stability. • Our model outperforms baselines by 43-46 % on small TSP instances. • We scale our method to 1400+ cities, beyond the reach of prior RL methods. The traveling salesman problem (TSP) is a classic NP-hard problem in combinatorial optimization with extensive practical applications. In this paper, we present an improved Actor-Critic architecture incorporating Proximal Policy Optimization (PPO) to effectively solve TSP. We introduce adaptive temperature scheduling, comprehensive state representation, and layer normalization to enhance learning stability. Experimental results demonstrate our Improved Actor-Critic approach achieves significant improvements ranging from 8.7 % to 55.9 % for different problem sizes compared to established reinforcement learning baselines including Q-Learning, SARSA, Double Q-Learning, Actor-Critic with Experience Replay (ACER), and Trust Region Policy Optimization (TRPO), with particularly strong performance on smaller instances between 20 to 100 cities. When testing on standard TSPLIB benchmarks, our method shows consistent advantages of 12 % to 33 % compared to classical approaches While tabular methods become computationally infeasible beyond 250 cities due to memory constraints, our approach maintains high solution quality for problems up to 1432 cities on our experimental setup (Intel® Core™i9-10900X CPU @ 3.70GHz × 20 with four NVIDIA Quadro RTX 5000 GPUs). Our ablation studies confirm the importance of each component in our proposed architecture, in which the improved state representation provides the most significant contribution to our model performance. This research significantly advances reinforcement learning approaches to combinatorial optimization, with practical implications for logistics, telecommunications, and manufacturing. The developed source code is available at: https://github.com/LetuQingge/TSP_Environment .

The rapid advancements in sequencing technologies have greatly increased access to genomic data stored in public databases. This has raised significant privacy and security concerns. This review emphasizes the importance of protecting genomic data by analyzing vulnerabilities in current storage and sharing practices. It examines the risks genetic databases face from cyber-attacks and internal breaches, focusing especially on advanced AI-driven threats and quantum computing vulnerabilities. The review explores machine learning methods designed to secure data. It highlights algorithms that prioritize privacy while maintaining data confidentiality, such as differential privacy, federated learning, and synthetic data generation using Generative Adversarial Networks (GANs). Findings demonstrate progress in mitigating common privacy breaches like re-identification and inference attacks. However, persistent vulnerabilities remain, particularly to emerging threats such as model inversion and membership inference attacks. The review advocates an integrated approach combining robust legislative frameworks with advanced technology to address genomic privacy challenges. It calls for intensified research efforts to safeguard genomic information. In particular, there is an urgent need to adopt quantum-resistant cryptographic methods, including lattice-based encryption and blockchain-integrated security frameworks. The paper emphasizes the necessity for genomics researchers to prioritize data privacy and security. This ensures responsible handling of genomic information in research.

The COVID-19 pandemic has underscored the need for effective diagnostic tools, particularly in resource-limited settings. While RT-PCR and CT scans are standard, their limitations drive the need for advanced techniques. This study leverages Convolutional Neural Networks, Knowledge Distillation, Ensemble Learning, and Federated Learning to develop robust, privacy-preserving models for COVID-19 detection from CT scans. We propose two federated learning strategies to simplify deep learning models for use in clinical environments with limited computational resources. The first strategy uses knowledge distillation from a complex model to a simplified model shared across a federated network. The second allows each hospital to distill knowledge to its simplified model, later combined into a global model via ensemble learning. Our methods, AFKD and IKDEFL, outperform traditional federated learning approaches such as FedAvg and FedAdam. AFKD, paired with the COVID-CNN model, achieves 91%-95% accuracy on IID (Independent and Identically Distributed) datasets and 70%-89% on non-IID datasets. IKDEFL further improves performance, with 92%-95% accuracy on IID datasets and 76%-88% on non-IID datasets. These approaches provide promising solutions for enhancing COVID-19 detection in federated learning.

In the realm of precision medicine, the prediction of phenotypic traits from genotype data plays a pivotal role in understanding disease susceptibility, drug response, and overall health outcomes. Machine learning models trained on genotype data offer promising avenues for phenotype prediction. However, the black-box nature of these models often hinders interpretabil-ity, raising concerns about their reliability and trustworthiness in clinical decision-making. This paper explores the application of SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) on our existing CNN model to identify important features from genotype that contribute to phenotype prediction in yeast data. Both techniques on machine learning models include improving model transparency and interpretability by providing insights into prediction mechanisms. LIME focuses on local interpretability, explaining individual predictions tailored to specific instances swiftly, while SHAP computes feature importance using Shapley values, offering a comprehensive understanding of feature contributions. Our dataset contains 4,390 samples, each with 28,220 features in genotype which are used in the training, testing and validation of our existing CNN model. In our analysis of 20 phenotypes, we identify the number of relevant important features using SHAP, LIME methods, and the intersection which not only improves the Mean Squared Error (MSE) values compared to using all the features in the genotype data, but also the model training time.

Mural restoration holds significant importance for various reasons from cultural, historical, and artistic aspects. Although significant progress has been made in natural image restoration in recent years, mural image restoration remains a challenging problem due to its differences in structure, color, and style from natural images. Motivated by the structural and color distribution characteristics of mural images, we propose a novel two-stage edge-guided mural in-painting model that perceives global and local information through a Transformer-Multi-scale CNN block (TMC) and learns informative color cues by introducing an extra YUV color space. Extensive experiments on Mural and CelebA-HQ datasets show that our proposed approach outperforms the other state-of-the-art baselines in both qualitative and quantitative evaluations.

Abstract Low‐light images are captured in environments with minimal lighting, such as nighttime or underwater conditions. These images often suffer from issues like low brightness, poor contrast, lack of detail, and overall darkness, significantly impairing human visual perception and subsequent high‐level visual tasks. Enhancing low‐light images holds great practical significance. Among the various existing methods for Low‐Light Image Enhancement (LLIE), those based on the Retinex theory have gained significant attention. However, despite considerable efforts in prior research, the challenge of Retinex decomposition remains unresolved. In this study, an LLIE network based on the Retinex theory is proposed, which addresses these challenges by integrating attention mechanisms and a U‐Net‐like architecture. The proposed model comprises three modules: the Decomposition module (DECM), the Reflectance Recovery module (REFM), and the Illumination Enhancement module (ILEM). Its objective is to decompose low‐light images based on the Retinex theory and enhance the decomposed reflectance and illumination maps using attention mechanisms and a U‐Net‐like architecture. We conducted extensive experiments on several widely used public datasets. The qualitative results demonstrate that the approach produces enhanced images with superior visual quality compared to the existing methods on all test datasets, especially for some extremely dark images. Furthermore, the quantitative evaluation results based on metrics PSNR, SSIM, LPIPS, BRISQUE, and MUSIQ show the proposed model achieves superior performance, with PSNR and BRISQUE significantly outperforming the baseline approaches, where (PSNR, mean BRISQUE) values of the proposed method and the second best results are (17.14, 17.72) and (16.44, 19.65). Additionally, further experimental results such as ablation studies indicate the effectiveness of the proposed model.

This paper introduces a novel parallel trajectory mechanism that combines Levenberg-Marquardt and Forward Accumulation Through Time algorithms to train a recurrent neural network controller in a closed-loop control system by distributing the calculation of trajectories across Central Processing Unit (CPU) cores/workers depending on the computing platforms, computing program languages, and software packages available. Without loss of generality, the recurrent neural network controller of a grid-connected converter for solar integration to a power system was selected as the benchmark test closed-loop control system. Two software packages were developed in Matlab and C++ to verify and demonstrate the efficiency of the proposed parallel training method. The training of the deep neural network controller was migrated from a single workstation to both cloud computing platforms and High-Performance Computing clusters. The training results show excellent speed-up performance, which significantly reduces the training time for a large number of trajectories with high sampling frequency, and further demonstrates the effectiveness and scalability of the proposed parallel mechanism.

Phenotype prediction refers to estimation or prediction of an organism's observable characteristics, or phenotypes, based on its genetic information or other relevant factors. The problem is important for several reasons. Phenotype prediction can contribute to a personalized medicine, prediction of disease risks, prognosis, and treatment response. In this paper, we develop quantitative phenotype regression prediction models to predict the values of phenotypes from various genotype features based on several machine learning and deep learning models, such as elastic net, a multilayer perceptron (MLP), convolutional neural network (CNN) and multi-task learning. The comparison results on yeast data show that our proposed elastic net and CNN model outperforms the existing models for 20 phenotype predictions.

This paper introduces a novel weight-dropout approach to train a neural network controller in real-time closed-loop control and to accelerate the embedded implementation for a solar inverter. The essence of the approach is to drop small-magnitude weights of neural network controllers during training with the goal of minimizing the required numbers of connections and guaranteeing the convergence of the neural network controllers. In order not to affect the convergence of neural network controllers, only non-diagonal elements of the neural network weight matrices were dropped. The dropout approach was incorporated into Levenberg-Marquardt and Forward Accumulation Through Time algorithms to train the neural network controller for trajectory tracking more efficiently. The Field Programmable Gate Array (FPGA) implementation on the Intel Cyclone V board shows significant improvement in terms of computation and resource requirements using the sparse weight matrices after dropout, which makes the neural network controller more suitable in an embedded environment.

Blind image quality assessment (BIQA) has made significant progress, but it remains a challenging problem due to the wide variation in image content and the diverse nature of distortions. To address these challenges and improve the adaptability of BIQA algorithms to different image contents and distortions, we propose a novel model that incorporates multiperspective consistency. Our approach introduces a multiperspective strategy to extract features from various viewpoints, enabling us to capture more beneficial cues from the image content. To map the extracted features to a scalar score, we employ a content‐aware hypernetwork architecture. Additionally, we integrate all perspectives by introducing a consistency supervision strategy, which leverages cues from each perspective and enforces a learning consistency constraint between them. To evaluate the effectiveness of our proposed approach, we conducted extensive experiments on five representative datasets. The results demonstrate that our method outperforms state‐of‐the‐art techniques on both authentic and synthetic distortion image databases. Furthermore, our approach exhibits excellent generalization ability. The source code is publicly available at https://github.com/gn-share/multi-perspective .

We are on the verge of a post-genomics era in which whole protein sequencing will be quickly carried out. Protein se-quencing plays an important role in identifying protein functions, analyzing protein-protein interactions, and characterizing post-translational modifications, etc. The protein sequencing problem is to determine the complete sequence of amino acids in proteins. De novo protein sequencing using top-down and bottom-up tandem mass spectrometry suffers from the problem of producing only partial sequences of target proteins, namely scaffold. In this paper, we explore the possibility of using deep learning techniques to perform the task of predicting amino acids in partially sequenced proteins by two phases. First, our methods involve querying the NCBI Protein Blast server to find closest matching homologous sequences to a scaffold as a training dataset. Second, we train several deep learning models based on a convolutional neural network and long short term memory to predict missing amino acids in the scaffold in the forward and reverse directions. We comprehensively evaluate our proposed methods on an alemtuzumab dataset and our results show that the proposed methods achieve high sequence coverage and high sequence accuracy with 100 % on the the light chain of alemtuzumab scaffold data.

Otsu's method is one of the most well‐known methods for automatic thresholding, which serves as an important algorithm category for image segmentation. However, it fails if the histogram is close to unimodal or has large intra‐class variances. To alleviate this limitation, improved Otsu's methods such as the valley emphasis method and weighted object variances method have been proposed, which still yield non‐optimal segmentation performance in some cases. In this study, a modified valley metric using second‐order derivative is proposed to improve the Otsu's algorithm. Experiments are firstly conducted on five typical test images whose histograms are unimodal, multimodal or have large intra‐class variances, and then expanded to a larger data set consisting of 22 cell images. The proposed algorithm is compared with original Otsu's method and existing improved algorithms. Four evaluation metrics including misclassification error, foreground recall, Dice similarity coefficient and Jaccard index are adopted to quantitatively measure the segmentation performance. Results show that the proposed algorithm achieves best segmentation results on both data sets quantitatively and qualitatively. The proposed algorithm adapts the Otsu's method to more image subtypes, indicating a wider application in automatic thresholding and image segmentation field.

The one-sided Exemplar Adjacency Number (EAN) is a known problem for computing the exemplar similarity between a generic linear genome${\mathcal G}$with gene duplications and an exemplar genome$H$(over the same set of$n$gene families). In this problem, we need to compute an exemplar genome$G$, which is a permutation obtained from${\mathcal G}$, such that the number of common adjacencies between$G$and$H$is maximized. Unfortunately, the problem is not only NP-hard but also NP-hard to approximate. In this paper, we approach the problem by relaxing the constraint such that a sub-permutation$G^{+}$obtained from${\mathcal G}$does not have to include all the gene families, but still needs to have a length at least$k$. Hence$G^{+}$is called apseudo-exemplargenome. Then, a slightly more general problem (One-sided EAN+) is defined: compute a pseudo-exemplar genome$G^{+}$from${\mathcal G}$such that the number of common adjacencies between$H$and$G^{+}$is maximized. Certainly One-sided EAN+ contains One-sided EAN as a special case; moreover, it presents some flexibility in designing algorithms. First, we relax and formulate the One-sided EAN+ problem as the maximum independent set (MIS) on a colored interval graph and hence reduce the appearance of each gene to at most two times. We show that this new relaxation is still NP-complete, though a simple factor-2 approximation algorithm can be designed; moreover, we also prove that the problem cannot be approximated within$2-\varepsilon$by a local search technique. We then show that this relaxed version is fixed-parameter tractable (FPT). Second, to ensure that each gene appears in$G^+$at most once, we use integer linear programming (ILP) to solve this problem. Finally, we implement our algorithm and compare it with the up-to-date software GREDU, with simulated signed and unsigned genomes. It turns out that our algorithm is more stable and can process genomes of length up to 12,000 for signed genomes (while GREDU can falter on such a large signed genome and it cannot handle unsigned genomes at all).