The multimodal multi-objective optimization problem (MMOP) involves multiple distinct components of the Pareto set (PS), which correspond to the same Pareto front (PF). Most existing multimodal multi-objective evolutionary algorithms (MMOEAs) face challenges in achieving both accuracy and timeliness in solving MMOPs. In this article, a two-stage clustering and independent competing based evolutionary algorithm (TSCICEA) is proposed for solving MMOPs. Firstly, in the early stage of population evolution, the K-means technique is used to coarsely cluster the population into multiple independent subpopulations, which aims to rapidly ascertain the approximate distribution structure of modalities and preliminarily locate their spatial positions. In the later stage, the DBSCAN technique is utilized to conduct fine clustering of the population, which aims to re-cluster misclustered individuals in the coarse clustering into new subpopulations, thus achieving a precise distinction between different modalities. Subsequently, to accurately identify the corresponding modality for each subpopulation, an independent competing strategy is proposed. This strategy can automatically assign different identification approaches for each modality based on their differences, and it simultaneously manages different modalities through parallel execution, thus significantly enhancing both the accuracy and timeliness of modality identification. To evaluate the effectiveness of TSCICEA, we compare it against state-of-the-art algorithms on benchmark test problems with 2–15 objectives and 4–15 variables. Experimental results demonstrate that TSCICEA achieves superior performance in terms of the IEDRX, IMST, and HV indicators

As a foundational technology for intelligent human-computer interaction, voice conversion (VC) seeks to transform speech from any source timbre into any target timbre. Traditional voice conversion methods based on Generative Adversarial Networks (GANs) encounter significant challenges in precisely encoding diverse speech elements and effectively synthesising these elements into natural-sounding converted speech. To overcome these limitations, we introduce Pureformer-VC, an encoder-decoder framework that utilizes Conformer blocks to build a disentangled encoder and employs Zipformer blocks to create a style transfer decoder. We adopt a variational decoupled training approach to isolate speech components using a Variational Autoencoder (VAE), complemented by triplet discriminative training to enhance the speaker’s discriminative capabilities. Furthermore, we incorporate the Attention Style Transfer Mechanism (ASTM) with Zipformer’s shared weights to improve the style transfer performance in the decoder. We conducted experiments on two multi-speaker datasets. The experimental results demonstrate that the proposed model achieves comparable subjective evaluation scores while significantly enhancing objective metrics compared to existing approaches in many-to-many and many-to-one VC scenarios.

Deep speech classification tasks, including keyword spotting and speaker verification, are vital in speech-based human-computer interaction. Recently, the security of these technologies has been revealed to be susceptible to backdoor attacks. Specifically, attackers use noisy disruption triggers and speech element triggers to produce poisoned speech samples that train models to become vulnerable. However, these methods typically create only a limited number of backdoors due to the inherent constraints of the trigger function. In this paper, we propose that speech backdoor attacks can strategically focus on speech elements such as timbre and emotion, leveraging the Speech Large Language Model (SLLM) to generate diverse triggers. Increasing the number of triggers may disproportionately elevate the poisoning rate, resulting in higher attack costs and a lower success rate per trigger. We introduce the Multiple Gradient Descent Algorithm (MGDA) as a mitigation strategy to address this challenge. The proposed attack is called the Speech Prompt Backdoor Attack (SPBA). Building on this foundation, we conducted attack experiments on two speech classification tasks, demonstrating that SPBA shows significant trigger effectiveness and achieves exceptional performance in attack metrics.

Modern deep learning models increasingly rely on third-party data processing, exposing vulnerabilities to backdoor attacks. Existing audio backdoor methods often compromise stealthiness by introducing perceptible modifications. This paper proposes Latent-space Frame-level Backdoor Attacks (LFBA), a novel framework that manipulates frame-level features in latent space to achieve imperceptible and effective backdoor injection. Our approach extracts and transforms frame-level features to subtly alter rhythmic patterns, such as compressing or expanding temporal segments, without modifying semantic content or speaker characteristics. Evaluations demonstrate excellent attack effectiveness while maintaining near-original audio quality. Our attack evades human perception and automated detection, maintaining robustness even after defensive fine-tuning. This work reveals critical risks in outsourced speech model training and establishes a new paradigm for stealthy, latent-space poisoning in speech-controlled systems.

In recent years at Tencent, we have observed that the use of multiple NoSQL databases for storing business data with diverse models has led to increased programming and deployment costs, as well as inefficient maintenance and underutilized resources. In this paper, we report X-Stor, a cloud-native NoSQL database system that supports multiple data models by extending different storage engines and efficiently managing them through a unified control plane. This design significantly reduces expenses and enables rapid expansion of new models, while seamlessly supporting their complete functionality through storage engine extensions. By consolidating multi-tenant services and data models on the same physical machines, X-Stor significantly enhances the utilization of cluster resources. Additionally, X-Stor introduces a standardized metric called Request Unit (RU) to measure tenant resource consumption for consumption-based pricing purposes. Leveraging this metric, we design RU-based resource management strategies and achieve efficient multi-tenant resource isolation and system load balancing. Currently, X-Stor manages a storage capacity of over 12PB for online operational data, including more than 100,000 tables with multiple data models. It handles 700 billion requests per day with a peak of 30 million requests per second. We evaluate the performance of X-Stor on popular benchmarks and production workloads. The results show that X-Stor performs well under diverse data models.

Hyperspectral images (HSI) have spectral variability, which leads to spectral dependence in adjacent and non-adjacent regions, and this dependence is essential for the classification of regions with mixed pixels. Current neural architecture search (NAS) methods have achieved significant advantages in HSI classification, but these methods cannot capture spectral dependence in non-adjacent regions because only use feedforward connections. Meanwhile, the cost of the search process in NAS is proportional to the scale of the search space, which limits the expansion of the search space. To address these issues, we propose a sparse-coding-inspired feedback neural architecture search (SCIF-NAS) method for HSI classification. Firstly, we view HSI samples as sequences and introduce a feedback mechanism in NAS to model the spectral dependence of non-adjacent regions to mitigate the effects of spectral variation. Secondly, we design several feedforward operations according to the characteristics of HSI, to form the search space together with feedback operations. Meanwhile, a sparse-coding-inspired NAS accelerated strategy is introduced to alleviate the search time burden caused by the expansion of search space. Thirdly, we integrate center loss with cross-entropy loss to construct a hybrid loss function that helps to obtain a better classification boundary. Finally, we conduct experiments on three popular HSI benchmarks, which show that SCIF-NAS outperforms the state-of-the-art methods in HSI classification.

Few-shot learning (FSL) has achieved promising performance in hyperspectral image classification (HSIC) with few labeled samples by designing a proper embedding feature extractor. However, the performance of embedding feature extractors relies on the design of efficient deep convolutional neural network architectures, which heavily depends on the expertise knowledge. Particularly, FSL requires extracting discriminative features effectively across different domains, which makes the construction even more challenging. In this paper, we propose a novel neural architecture search-based FSL model for HSI classification, called HCFSL-NAS. Three novel strategies are proposed in this work. First, a neural architecture search-based embedding feature extractor is developed to the FSL in HSIC, whose search space includes a group of proposed multi-scale convolutions with channel attention. Second, a multi-source learning framework is employed to aggregate abundant heterogeneous and homogeneous source data, which enables the powerful generalization of network to the HSIC with only few labeled samples. Finally, the pointwise-based cross-entropy loss and the pairwise-based adaptive sparse loss are jointly optimized to maximize inter-class distance and minimize the distance within a class simultaneously. Experimental results on four publicly hyperspectral data sets demonstrate that HCFSL-NAS outperforms both the exiting FSL methods and supervised learning methods for HSI classification with only few labeled samples. Code is available at: https://github.com/xh-captain/HCFSL-NAS.

RGB-D salient object detection (SOD) focuses on utilizing the complementary cues of RGB and depth modalities to detect and segment salient regions. However, many proposed methods train their models in a simple multi-modal manner, ignoring the differences between these two modalities in the contribution of salient detection. Furthermore, the quality of depth datasets varies significantly between individuals and is another important factor affecting model performance. To address the aforementioned issues, this article proposes a novel depth-guided fusion network framework (DGFNet) for the RGB-D SOD task. To avoid the influence of low-quality depth maps on RGB-D SOD, we design a depth map enhanced algorithm which jointly models salient detection and depth estimation to improve the quality of depth. Also, we propose a depth attention mechanism to encode valuable spatial information for SOD, which is then used in depth-guided fusion (DGF) module to guide the fusion of cross-modality features at each level. Extensive experiments on seven commonly tested datasets demonstrate that our DGFNet outperforms the 23 state-of-the-art RGB-D-based SOD methods.

Long-running containerized workloads (e.g., machine learning), which typically showtime-varyingpatterns, are increasingly prevailing in shared production clusters. To improve workload performance, current schedulers mainly focus on optimizingshort-termbenefits of cluster load balancing orinitial container placementon servers. However, this would inevitably bring manyinvalid migrations(i.e., containers are migrated back and forth among servers over a short time window), leading to significant service level objective (SLO) violations. This paper introducesTetris, amodel predictive control(MPC)-based container scheduling strategy to proactively migrate long-running workloads for cluster load balancing. Specifically, we first build a discrete-time dynamic model forlong-termoptimization of container scheduling. To solve such an optimization problem,Tetristhen employs two main components: (1) a container resource predictor, which leverages time-series analysis approaches to accurately predict the container resource consumption; (2) an MPC-based container scheduler that jointly optimizes the cluster load balancing and container migration costover a certain sliding time window. We implement and open source a prototype ofTetrisbased on K8s. Extensive prototype experiments and trace-driven simulations demonstrate thatTetriscan improve the cluster load balancing degree by up to 77.8% without incurring any SLO violations, compared to the state-of-the-art container scheduling strategies.

Hyperspectral image (HSI) classification based on neural architecture search (NAS) is a currently attractive frontier as it not only automatically searches complex neural network architecture, but also avoids professional knowledge and experience design, and alleviates the lacking of generalization ability as well when dealing with a new classification task. However, the existing HSI classification based on NAS has some drawbacks: 1) A huge number of training parameters and high calculations are inductive to over-fitting and high complexity. 2) Efficient operators are lacking in the search space which can distinguish spatial locations and spectral features in different bands. Furthermore, as the category samples in HSI data show a serious long-tail distribution phenomenon, HSI classification remains challenging. To address these issues, we propose a lightweight HSI classification model LMSS-NAS integrating multi-scale spectral-spatial attention. The main work includes three-fold: 1) In order to reduce the number of model parameters and promote spectral-spatial feature fusion, a new lightweight efficient search space is designed, which consists of three equivalent lightweight convolution operators with multiple receptive fields. 2) To fully use the spectral-spatial correlation of HSI, a cube-to-pixel classification framework is designed to mine the local spatial and spectral context. 3) Focal loss and label smoothing loss in computer vision tasks are jointly migrated to LMSS-NAS to improve the unbalanced samples’ classification and model robustness. Experimental results on four public hyperspectral data sets show that the proposed method can achieve competitive classification performance as well as low computational cost. Code is available at: https://github.com/xh-captain/LMSS-NAS.

Breast cancer diagnosis in multi-gigapixel whole slide images (WSIs) is an important task that highly relevant to cancer grading and prognosis. In recent years, many computer-aided diagnosis methods were proposed and achieved promising performance. However, they mostly suffer from heavy computational burden that becomes a significant barrier to clinical practice. Efficient solutions are urgently demanded but still less studied. In this paper, we propose a novel framework named EffiDiag for a fast and lightweight breast cancer diagnosis. To this end, a loss-modified U-net is developed at first to enable a fast suspected cancer Region Of Interest (ROI) localization. Therefore the subsequent patch-based classification, which commonly executes at the finest magnification hundreds of thousands times per WSI for cancer identification, could be carried out on these ROIs only rather than the whole WSI for speedup. Meanwhile, a super-efficient convolutional neural network (CNN) is devised to optimize the classification speed and resource consumption per classification. Experiments on the Camelyonl6 benchmark demonstrate, by integrating the two contributions into a well-established approach, 47x inference acceleration is obtained with limited accuracy drop, yet with much less resource consumption even compared to popular lightweight networks.

Compressive sensing magnetic resonance (MR) imaging is aimed at achieving high‐quality MR image reconstruction by undersampling K‐space data. It is crucial to explore prior information since compressive sensing MR imaging relies heavily on some prior assumptions, such as signal's sparse property. In this study, in order to explore the prior information fully, an improved MR image reconstruction model based on compressive sensing theory is proposed, named reference image MR imaging with adaptive tight frame. In the proposed model, an adaptive tight frame is involved to explore the sparse prior information adapt to MR images and the similarity prior information to the target image. Meanwhile, improved adaptive weighting parameters are used to trade off the sparsity between the regions with much similarity and that of little similarity. In addition, the smoothing‐based fast iterative shrinkage‐threshold algorithm is utilised to tackle the optimisation problem so as to speed up imaging. The experimental results demonstrate that the proposed MR image reconstruction method outperforms some state‐of‐the‐art methods in terms of quantitative results.

In bioinformatics, machine learning-based prediction of drug-target interaction (DTI) plays an important role in virtual screening of drug discovery. DTI prediction, which have been treated as a binary classification problem, depends on the concentration of two molecules, the interaction between two molecules, and other factors. The degree of affinity between a drug molecule (such as a drug compound) and a target molecule (such as a receptor or protein kinase) reflects how tightly the drug binds to a particular target and is quantified by the measurement which can reflect more detailed and specific information than binary relationship. In this study, we proposed an end-to-end model, named AttentionDTA, based on deep learning, which associates attention mechanism to predict the binding affinity of DTI. The novelty in this work is to use attentional mechanisms to consider which subsequences in a protein are more important for a drug and which subsequences in a drug are more important for a protein when predicting its affinity. So that the representational ability of the model is stronger. The model uses one-dimensional Convolution Neural Networks (1D-CNNs) to extract the abstract information of drug and protein, and makes the drug and protein representations mutually adapt through the attention mechanisms. We evaluate our model on two established drug-target affinity benchmark datasets, Davis and KIBA. The model outperforms DeepDTA, a state-of-the-art deep learning method for drug-target binding affinity prediction, with better Mean Squared Error (MSE), Concordance Index (CI), rm2, and Area Under Precision Recall Curve (AUPR). Our results show that the attention-based model can effectively extract effective representations by calculating the weight of the representation between the drug and the protein. Finally, we visualize the attention weight. It proves our model can obtain the information of binding sites.

Object detection in computer vision area has been extensively studied and making tremendous progress in recent years using deep learning methods. However, due to the heavy computation required in deep learning based algorithms, it is hard to run these models on embedded systems, which have limited computing capabilities. In response to this situation, we combine the idea of MobileNet to improve RetinaNet, design a lightweight deep learning model MRNet which not only suitable for embedded systems but also achieve high accuracy. The model was demonstrated in the MMSP 2019 Embedded Deep Learning Object Detection Model Competition. In the comprehensive evaluation of the model size, computational complexity and running speed on TX2. our proposed method won the third place in the competition.

Salient object detection is a fundamental problem and has been received a great deal of attention in computer vision. Recently, deep learning model became a powerful tool for image feature extraction. In this study, the authors propose a multi‐scale deep neural network (MSDNN) for salient object detection. The proposed model first extracts global high‐level features and context information over the whole source image with the recurrent convolutional neural network. Then several stacked deconvolutional layers are adopted to get the multi‐scale feature representation and obtain a series of saliency maps. Finally, the authors investigate a fusion convolution module to build a final pixel level saliency map. The proposed model is extensively evaluated on six salient object detection benchmark datasets. Results show that the authors’ deep model significantly outperforms other 12 state‐of‐the‐art approaches.

Multidimensional linear phase perfect reconstruction filter bank (MDLPPRFB) can be designed and implemented via lattice structure. The lattice structure for the MDLPPRFB with filter support N(MΞ) has been published by Muramatsu , where M is the decimation matrix, Ξ is a positive integer diagonal matrix, and N(N) denotes the set of integer vectors in the fundamental parallelepiped of the matrix N. Obviously, if Ξ is chosen to be other positive diagonal matrices instead of only positive integer ones, the corresponding lattice structure would provide more choices of filter banks, offering better trade-off between filter support and filter performance. We call such resulted filter bank as generalized-support MDLPPRFB (GSMDLPPRFB). The lattice structure for GSMDLPPRFB, however, cannot be designed by simply generalizing the process that Muramatsu employed. Furthermore, the related theories to assist the design also become different from those used by Muramatsu . Such issues will be addressed in this paper. To guide the design of GSMDLPPRFB, the necessary and sufficient conditions are established for a generalized-support multidimensional filter bank to be linear-phase. To determine the cases we can find a GSMDLPPRFB, the necessary conditions about the existence of it are proposed to be related with filter support and symmetry polarity (i.e., the number of symmetric filters ns and antisymmetric filters na). Based on a process (different from the one Muramatsu used) that combines several polyphase matrices to construct the starting block, one of the core building blocks of lattice structure, the lattice structure for GSMDLPPRFB is developed and shown to be minimal. Additionally, the result in this paper includes Muramatsu's as a special case.

Semantic similarity between ontological concepts plays an important role in service discovery and composition. In this paper, using the ontological relation, a novel intuitionistic fuzzy set model is proposed to interpret concepts on the ontology with three intuitionistic fuzzy sets of different weights, and an effective similarity measure of which is selected to calculate the concept similarity. The model successfully translates all the semantic information of a concept into mathematical expressions of intuitionistic fuzzy sets, thus concept similarity is easily calculated. It is designed to handle not only simple ontologies where only atomic concepts are presented, but also complex ones where concepts could inherit from multiple concepts and have semantic relations other than is-a with other concepts. The experimental result has shown that the model is effective, and it's easy to implement.

This paper focuses on the construction of multidimensional biorthogonal multiwavelets and the perfect reconstruction multifilter banks. Based on the Hermite-Neville filter, two lifting structures have been proposed and systematically investigated, and a general design framework has been developed for building biorthogonal multiwavelets and Hermite interpolation filter banks with any multiplicity for any lattice in any dimension with any number of primal and dual vanishing moments. The construction is an important generalization of the Neville-based lifting scheme and inherits all of the advantages of lifting schemes such as fast transform, in-place computation and integer-to-integer transforms. Our multiwavelet systems preserve most of the desirable properties for applications, such as interpolating, short support, symmetry, and high vanishing moments.