Learning Iteration Research Articles

Trust driven On-Demand scheme for client deployment in Federated Learning

Containerization technology plays a crucial role in Federated Learning (FL) setups, expanding the pool of potential clients and ensuring the availability of specific subsets for each learning iteration. However, doubts arise about the trustworthiness of devices deployed as clients in FL scenarios, especially when container deployment processes are involved. Addressing these challenges is important, particularly in managing potentially malicious clients capable of disrupting the learning process or compromising the entire model. In our research, we are motivated to integrate a trust element into the client selection and model deployment processes within our system architecture. This is a feature lacking in the initial client selection and deployment mechanism of the On-Demand architecture. We introduce a trust mechanism, named “Trusted-On-Demand-FL”, which establishes a relationship of trust between the server and the pool of eligible clients. Utilizing Docker in our deployment strategy enables us to monitor and validate participant actions effectively, ensuring strict adherence to agreed-upon protocols while strengthening defenses against unauthorized data access or tampering. Our simulations rely on continuous user behavior datasets, deploying an optimization model powered by a genetic algorithm to efficiently select clients for participation. By assigning trust values to individual clients and dynamically adjusting these values, combined with penalizing malicious clients through decreased trust scores, our proposed framework identifies and isolates harmful clients. This approach not only reduces disruptions to regular rounds but also minimizes instances of round dismissal, Consequently enhancing both system stability and security.

A study of the role of data and model uncertainty in active learning

Uncertainty-based active learning strategies have demonstrated significant superiority in small data research of materials domain. This study explores the effects of model uncertainty and data uncertainty separately on the performance of active learning strategies, specifically focusing on the number of iterations required to identify the optimal samples. For model uncertainty, three kinds of acquisition functions are compared, including predicted value strategy (PV), ranking of predicted value strategy (PR) and expected improvement strategy (EI). Among these, the active learning model utilizing PR requires the fewest average iterations (1.75). For data uncertainty, we evaluate the iterations of active learning by Gaussian process models that incorporate the uncertainty of the observations and noise samples that takes account into the uncertainty of the input features respectively. The results indicate that the active learning iterations of the three strategies converge to similar at the optimal weighting when the uncertainty of the observations is considered in the model (EI for 1.75, PV for 1.21 and PR for 1.18). In contrast, incorporating noise samples into the augmented dataset after the original samples would severely deteriorate the efficiency of active learning recommendations. Our findings aim to offer guidance for exploring more favorable acquisition functions and methods for active learning strategies.

A comparative study of vehicle frame load spectrum extraction using SA-1DCNN deep learning and virtual iteration

A comparative study of vehicle frame load spectrum extraction using SA-1DCNN deep learning and virtual iteration

ANS-SCMC: A matrix completion method based on adaptive neighbourhood similarity and sparse constraints for predicting microbe-disease associations.

The use of matrix completion methods to predict the association between microbes and diseases can effectively improve treatment efficiency. However, the similarity measures used in the existing methods are often influenced by various factors such as neighbourhood size, choice of similarity metric, or multiple parameters for similarity fusion, making it challenging. Additionally, matrix completion is currently limited by the sparsity of the initial association matrix, which restricts its predictive performance. To address these problems, we propose a matrix completion method based on adaptive neighbourhood similarity and sparse constraints (ANS-SCMC) for predict microbe-disease potential associations. Adaptive neighbourhood similarity learning dynamically uses the decomposition results as effective information for the next learning iteration by simultaneously performing local manifold structure learning and decomposition. This approach effectively preserves fine local structure information and avoids the influence of weight parameters directly involved in similarity measurement. Additionally, the sparse constraint-based matrix completion approach can better handle the sparsity challenge in the association matrix. Finally, the algorithm we proposed has achieved significantly higher predictive performance in the validation compared to several commonly used prediction methods proposed to date. Furthermore, in the case study, the prediction algorithm achieved an accuracy of up to 80% for the top 10 microbes associated with type 1 diabetes and 100% for Crohn's disease respectively.

The bounded intelligence of AI: Superficiality and deceivability

Artificial Intelligence (AI) is typically designed to replace or augment human beings in making and executing decisions. In the domain of organizational management, AI has already found wide applications in many areas such as HR functions, marketing campaigns, competitive analysis, and strategy formulations, exerting increasing impacts on a host of executives and management professionals. Accurate and timely replicating or capturing the know-how of human decisions and actions becomes imperative, as AI has to be, at least as intelligent as, or even more intelligent than, those human actors whose data are being fed to AI during its training and learning iterations. Just as human beings are subject to bounded rationality due to their limited capacities in information processing, AI, too, is subject to bounded intelligence, in the sense that it could appear not as smart as expected, due to its limited capacities in replicating and capturing human expertise, to say nothing of augmenting or extending it. Such bounded intelligence of AI hinges on two underlying factors: superficiality and deceivability. Superficiality refers to the inability of AI to fully replicate human expertise during its learning process due to natural or technical barriers such as imperfect capturability, relational embeddedness, and time sensitivity. Deceivability refers to the inability of AI to accurately capture human expertise during its learning process due to human-actor-created barriers such as passive evasion, deliberate manipulation, and malicious sabotage. Understanding such boundedness of AI and the specific mechanisms of superficiality and deceivability will help us better appreciate the usefulness and limitations of AI and take measures to both remedy and utilize the bounded intelligence of AI.

An improved YOLOv8 model enhanced with detail and global features for underwater object detection

Underwater object detection is significant for the practical research of mastering existing marine biological resources. In response to the challenges posed by complex underwater environments such as water scattering and variations in object scales, researchers have developed YOLOv8 for object detection, driven by the rising popularity and iteration of deep learning. Building upon this model, we propose an enhanced underwater object detection model named YOLOv8-DGF. Firstly, we replace the convolutional layers of Spatial Pyramid Pooling Fusion (SPPF) with Invertible Neural Networks to further augment the fusion capacity of detailed features, facilitating the preservation of pivotal information while mitigating the impact of noise. Additionally, we introduce a global attention mechanism into Convolution to Fully Connected (C2f), which weights the input features, thereby emphasizing or suppressing feature information from different locations. Through our ‘Detail to Global’ strategy, the model achieved mAP@0.5 scores of 87.7% and 84.8% on the RUOD and URPC2020 datasets, respectively, with improved processing speed. Extensive ablation experiments on the Pascal VOC dataset demonstrate that YOLOv8-DGF outperforms other methods, achieving the best overall performance.

Machine learning-guided multi-site combinatorial mutagenesis enhances the thermostability of pectin lyase

Enhancing the thermostability of enzymes is crucial for industrial applications. Methods such as directed evolution are often limited by the huge sequence space and combinatorial explosion, making it difficult to obtain optimal mutants. In recent years, machine learning (ML)-guided protein engineering has become an attractive tool because of its ability to comprehensively explore the sequence space of enzymes and discover superior mutants. This study employed ML to perform combinatorial mutation design on the pectin lyase PMGL-Ba from Bacillus licheniformis, aiming to improve its thermostability. First, 18 single-point mutants with enhanced thermostability were identified through semi-rational design. Subsequently, the initial library containing a small number of low-order mutants was utilized to construct an ML model to explore the combinatorial sequence space (theoretically 196,608 mutants) of single-point mutants. The results showed that the ML-predicted second library was successfully enriched with highly thermostable combinatorial mutants. After one iteration of learning, the best-performing combinatorial mutant in the third library, P36, showed a 67-fold and 39-fold increase in half-life at 75 °C and 80 °C, respectively, as well as a 2.1-fold increase in activity. Structural analysis and molecular dynamics simulations provided insights into the improved performance of the engineered enzyme.

Optimizing robotic arm control using deep Q-learning and artificial neural networks through demonstration-based methodologies: A case study of dynamic and static conditions

This paper uses robot programming techniques, such as Deep Q Network, Artificial Neural Network, and Artificial Deep Q Network, to address challenges related to controlling robotic arms through demonstration learning. Static and dynamic states of the subjects were the subjects of experiments. Each method's classification accuracy process success values and experimental condition combination were evaluated. The DQN method demonstrated favourable classification accuracy outcomes, achieving an Accuracy value of 0.64 for the fixed dice and 0.52 for the moving dice. The Response value was 0.51 for the fixed dice and 0.41 for the moving dice, indicating a moderate level. The ANN method demonstrated lower accuracy, with Accuracy values of 0.59 and 0.56 and Response values of 0.61 and 0.58, respectively. The ADQN method demonstrated superior outcomes, with Accuracy values of 0.66 and 0.59 and Response values of 0.67 and 0.61. During the initial learning iterations, ADQN demonstrated the highest success rate at 33.67 %, whereas DQN and ANN achieved 28.39 % and 20.13 % success rates, respectively. As the number of iterations increased, all methods demonstrated improvement in their results. ADQN maintained a high success rate of 97.59 %, while DQN and ANN attained 82.16 % and 88.66 %, respectively. As the number of iterations increases, the results of all methods improve, but the success rate of the Artificial Deep Q Network remains high. As the number of iterations increases, both Deep Q Network and Artificial Neural Network demonstrate the potential to achieve good results. Overall, the findings support the efficacy of robot programming techniques that incorporate demonstration learning. The Artificial Deep Q Network is the most successful and fast-converging method suitable for various robot control tasks. These findings provide a foundation for future research and large-scale, comprehensive learning applications for complex rot control.

Real‐time fire and smoke detection with transfer learning based on cloud‐edge collaborative architecture

AbstractRecent years have seen increased interest in object detection‐based applications for fire detection in digital images and videos from edge devices. The environment's complexity and variability often lead to interference from factors such as fire and smoke characteristics, background noise, and camera settings like angle, sharpness, and exposure, which hampers the effectiveness of fire detection applications. Limited picture data for fire and smoke scenes further challenges model accuracy and robustness, resulting in high false detection and leakage rates. To address the need for efficient detection and adaptability to various environments, this paper focuses on (1) proposing a cloud‐edge collaborative architecture for real‐time fire and smoke detection, incorporating an iterative transfer learning strategy based on user feedback to enhance adaptability; (2) improving the detection capabilities of the base model YOLOv8 by enhancing the data augmentation method and introducing the coordinate attention mechanism to improve global feature extraction. The improved algorithm shows a 2‐point accuracy increase. After three iterations of transfer learning in the production environment, accuracy improves from 93.3% to 96.4%, and mAP0.5:0.95 increases by nearly 5 points. This program effectively addresses false detection issues in fire and smoke detection systems, demonstrating practical applicability.

Key factor screening in mouse NASH model using single-cell sequencing combined with machine learning

AimsTo identify and analyze genes closely related to the progression of nonalcoholic steatohepatitis (NASH) by employing a combination of single-cell RNA sequencing and machine-learning algorithms. Main methodsSingle-cell RNA sequencing (scRNA-seq) analysis was performed to find the cell population with the most significant differences between the Chow and NASH groups. This approach was used to validate the developmental trajectory of this cell population and investigate changes in cellular communication and important signaling pathways among these cells. Subsequently, high dimensional Weighted Gene Co-expression Network Analysis (hdWGCNA) was used to find the key modules in NASH. Machine learning analyses were performed to further identify core genes. Deep learning techniques were applied to elucidate the correlation between core genes and immune cells. The accuracy of this correlation was further confirmed using deep learning techniques, specifically Convolutional Neural Networks. Key findingsBy comparing scRNA-seq data between the Chow and NASH groups, we have observed a notable distinction existing in the Kupffer cell population. Signaling interactions between hepatic macrophages and other cells were significantly heightened in the NASH group. Through subsequent analysis of macrophage subtypes and key modules, we identified 150 genes tightly associated with NASH. Finally, we highlighted the 16 most significant core genes using multiple iterations of machine learning. Furthermore, we pointed out the close relationship between core genes and immune cells. SignificancesUsing scRNA-seq analysis and machine learning, we can distinguish NASH-related genes from large genetic datasets, providing theoretical support in finding potential targets for the development of novel therapies.

Reducing annotating load: Active learning with synthetic images in surgical instrument segmentation

Accurate instrument segmentation in the endoscopic vision of minimally invasive surgery is challenging due to complex instruments and environments. Deep learning techniques have shown competitive performance in recent years. However, deep learning usually requires a large amount of labeled data to achieve accurate prediction, which poses a significant workload. To alleviate this workload, we propose an active learning-based framework to generate synthetic images for efficient neural network training. In each active learning iteration, a small number of informative unlabeled images are first queried by active learning and manually labeled. Next, synthetic images are generated based on these selected images. The instruments and backgrounds are cropped out and randomly combined with blending and fusion near the boundary. The proposed method leverages the advantage of both active learning and synthetic images. The effectiveness of the proposed method is validated on two sinus surgery datasets and one intraabdominal surgery dataset. The results indicate a considerable performance improvement, especially when the size of the annotated dataset is small. All the code is open-sourced at: https://github.com/HaonanPeng/active_syn_generator.

Self-paced method for transfer partial label learning

In partial label learning (PLL) problem, each training sample corresponds to a group of candidate labels, in which only one label is the ground-truth label (correct label). Almost all the existing PLL algorithms attempt to eliminate the ambiguity of the candidate label sets by treating all labels indiscriminately. However, this kind of approach lacks the consideration of the complexity between the labels and the instances in the training process. Encouraged by the extensive researches of the self-paced learning (SPL) and transfer learning (TL) in various fields, this paper introduces SPL and TL together to address the PLL problem and proposes a new SPL framework, which is called self-paced method for transfer partial label learning (SPTPLL). The proposed model utilizes transfer learning model to share the parameters and regularization terms of the Support Vector Machine (SVM), which can transfer knowledge from the source task to the target task. Additionally, we implement the self-paced learning scheme by choosing a suitable self-paced function to enhance the robustness of the proposed model. In the process of learning iteration, the priority of training examples with their candidate labels is ranked through self-paced learning to control the learning process. Finally, we demonstrate the superior performance of the proposed method through a large number of experiments compared with state-of-the-art baseline methods.

Recognition of Plastic Film in Terrain-Fragmented Areas Based on Drone Visible Light Images

In order to meet the growing demand for food and achieve food security development goals, contemporary agriculture increasingly depends on plastic coverings such as agricultural plastic films. The remote sensing-based identification of these plastic films has gradually become a necessary tool for agricultural production management and soil pollution prevention. Addressing the challenges posed by the complex terrain and fragmented land parcels in karst mountainous regions, as well as the frequent presence of cloudy and foggy weather conditions, the extraction efficacy of mulching films is compromised. This study utilized a DJI Mavic 2 Pro UAV to capture visible light images in an area with complex terrain features such as peaks and valleys. A plastic film sample dataset was constructed, and the U-Net deep learning model parameters integrated into ArcGIS Pro were continuously modified and optimized to achieve precise plastic film identification. The results are as follows: (1) Sample quantity significantly affects recognition performance. When the sample size is 800, the accuracy of plastic film extraction notably improves, with area accuracy reaching 91%, a patch quantity accuracy of 96.38%, and an IOU and F1-score of 85.89% and 94.20%, respectively, compared to the precision achieved with a sample size of 300; (2) Different learning rates, batch sizes, and iteration numbers have a certain impact on the training effectiveness of the U-Net model. The most suitable model parameters improved the training effectiveness, with the highest training accuracy achieved at a learning rate of 0.001, a batch size of 10, and 25 iterations; (3) Comparative experiments with the Support Vector Machine (SVM) model validate the suitability of U-Net model parameters and sample datasets for precise identification in rugged terrains with fragmented spatial distribution, particularly in karst mountainous regions. This underscores the applicability of the U-Net model in recognizing plastic film coverings in karst mountainous regions, offering valuable insights for agricultural environmental health assessment and green planting management in farmlands.

A Novel Accelerated Multistage Learning Control Mechanism via Virtual Performance Reduction.

This study uses a multistage learning mechanism concept to investigate the accelerated learning control for stochastic systems. In this mechanism, the learning iterations are divided into successive stages, with each stage comprising several iterations. The learning gain is constant in each stage to accelerate the learning process and decreases it from one stage to another to eliminate the noise effect asymptotically. The critical issue is determining the switching iteration when a new stage starts. This study resolves this issue by calculating a virtual performance index of the mean-squared input error and its estimated upper bound. Specifically, the ideal, practical, and improved multistage learning control schemes are proposed to determine the switching iteration and generate the learning gain sequence. The ideal scheme achieves the best performance at the cost of a large computation burden, and the practical scheme saves computation cost, but the performance is not excellent. The improved scheme significantly approximates the best performance by introducing additional stretching parameters to the performance index. Illustrative simulations are provided to verify the theoretical results.

ProcessGallery: Contrasting Early and Late Iterations for Design Principle Learning

Traditional design galleries enable users to search for examples based on surface attributes (e.g., color or style), and largely obscure underlying principles (e.g., hierarchy or readability). We conducted three studies to explore how galleries could be constructed to help novices learn key design principles. Study 1 revealed that novices gain perspective by observing how designs evolve throughout a process. Study 2 found that novices are better at identifying design issues when viewing iterations that show improvements for just one principle at a time, rather than multiple. Building on these insights, we created ProcessGallery, a tool that enables users to browse contrasting pairs of early-and-late iterations of designs that highlight key improvements organized by design principles. In Study 3, a within-subjects experiment, sixteen participants iterated on a seed design after viewing examples in ProcessGallery versus a traditional gallery. Using ProcessGallery, participants found more appropriate examples, assessed designs better, and preferred ProcessGallery for learning compared to a traditional gallery.

Design of high performance Cu-Ni-Si alloys via a multiobjective strategy based on machine learning

Herein, the ultimate tensile strength and electrical conductivity of precipitation-strengthened Cu-Ni-Si alloys were simultaneously improved by utilizing a machine learning-based multiobjective design strategy. The multiobjective design strategy consists of five main steps: creating the initial dataset, generating alloy features, screening key alloy features, modeling and inversely designing, and experimental iteration. Of particular note is the constraint placed on the initial composition-properties dataset, considering the rules governing the addition of Co. This constraint ensures that the dataset adheres to the required specifications. To evaluate the optimized degree of the inverse design composition, a joint expectation improvement function was employed. This function effectively integrates the ultimate tensile strength and electrical conductivity. Through a process of five mutually reinforcing iterations of machine learning and experimentation, the combined property of the designed alloy surpasses the Pareto frontier initially formed by the collected data. Microstructure analysis further confirmed the significant precipitation strengthening effects achieved in the optimized alloy.

FedCGS: Leverage Feature Compression and Conditional GAN for Privacy Protection and Efficient Communication in Federated Learning

Federated learning is a distributed machine learning technique that can train models from decentralized data, collaboratively building central models and protecting data privacy by passing partial model updates rather than raw data in each iteration of model learning. However, recent work has shown that the information during the exchange between the server and the client is still vulnerable to gradient-based privacy attacks, and traditional privacy protection methods can incur significant costs (such as communication consumption and performance evaluation). Therefore, we need an algorithm that can protect data privacy while reducing communication costs and performance losses. Here, we propose a federated learning framework FedCGS that leverages conditional generation adversarial networks and feature compression in order to achieve privacy protection while maintaining competitive model performance and effectively improving communication efficiency. We divided the local network of each client into feature extractor and public classifier, and kept the extractor local to protect privacy, and sent classifier and client generator to the server for aggregation and update through singular value decomposition, so as to improve the performance of the local network and efficient communication. A lot of experimental data prove that FedCGS can improve the performance of local model better than traditional federation learning, and improve the communication efficiency better than common conditional generation adversarial network. Therefore, we believe that FedCGS has better federal learning potential

LSTKC: Long Short-Term Knowledge Consolidation for Lifelong Person Re-identification

Lifelong person re-identification (LReID) aims to train a unified model from diverse data sources step by step. The severe domain gaps between different training steps result in catastrophic forgetting in LReID, and existing methods mainly rely on data replay and knowledge distillation techniques to handle this issue. However, the former solution needs to store historical exemplars which inevitably impedes data privacy. The existing knowledge distillation-based models usually retain all the knowledge of the learned old models without any selections, which will inevitably include erroneous and detrimental knowledge that severely impacts the learning performance of the new model. To address these issues, we propose an exemplar-free LReID method named LongShort Term Knowledge Consolidation (LSTKC) that contains a Rectification-based Short-Term Knowledge Transfer module (R-STKT) and an Estimation-based Long-Term Knowledge Consolidation module (E-LTKC). For each learning iteration within one training step, R-STKT aims to filter and rectify the erroneous knowledge contained in the old model and transfer the rectified knowledge to facilitate the short-term learning of the new model. Meanwhile, once one training step is finished, E-LTKC proposes to further consolidate the learned long-term knowledge via adaptively fusing the parameters of models from different steps. Consequently, experimental results show that our LSTKC exceeds the state-of-the-art methods by 6.3%/9.4% and 7.9%/4.5%, 6.4%/8.0% and 9.0%/5.5% average mAP/R@1 on seen and unseen domains under two different training orders of the challenging LReID benchmark respectively.

Investigating Suitable Combinations of Dynamic Models and Control Techniques for Offline Reinforcement Learning Based Navigation: Application of Universal Omni-Wheeled Robots

Abstract Omnidirectional locomotion provides wheeled mobile robots (WMR) with better maneuverability and flexibility, which enhances their energy efficiency and dexterity. Universal omni-wheels are one of the best categories of wheels that can be used to develop a WMR (Amarasiri et al., 2022, “Robust Dynamic Modeling and Trajectory Tracking Controller of a Universal Omni-Wheeled Mobile Robot,” ASME Letters Dyn. Sys. Control., 2(4), p. 040902. 10.1115/1.4055690). We study dynamic modeling and controllers for mobile robots to train in a reinforcement learning (RL)-based navigation algorithm. RL tasks require copious amounts of learning iteration episodes, which makes training very time consuming. The choice of dynamic model and controller has a significant impact on training time. In this paper, we compare a traditional Kane’s equations model to a non-holonomic canonical momenta model (Barhorst, 2019, “Generalized Momenta in Constrained Non-Holonomic Systems—Another Perspective on the Canonical Equations of Motion,” Int. J. Non-Linear Mech., 113, pp. 128–145.). We implemented four controllers: proportional integral derivative, linear quadratic regulator with integral action, pole placement, and a full nonlinear sliding mode controller. We summarize the pros and cons of each of the modeling techniques, and control laws implemented. The outcomes of our analysis will improve RL training time for path generation in unstructured environments.

Active Crowdsourcing for Multilabel Annotation.

Multilabel annotation is a critical step to generate training sets when learning classification models in various application domains, but asking domain experts to provide labels is usually time-consuming and expensive, which cannot meet the current requirement of the fast evolution of the models in the big data era. Although crowdsourcing provides a fast solution to acquire labels for multilabel learning, it faces the risk of high data acquisition cost and low label quality. This article proposes a novel one-coin label-dependent active crowdsourcing (OCLDAC) method to iteratively query noisy labels from crowd workers and learn multilabel classification models. In each iteration of active learning, integrated labels of instances are first inferred by a novel one-coin label-dependent model, which utilizes a mixture of multiple independent Bernoulli distributions to explore and exploit correlations among the labels to increase the accuracy of truth inference. Then, instances, labels, and workers are selected according to the novel strategies that incorporate the distribution of noisy labels, the prediction probability of learning models, label correlations, and the reliability of crowd workers. Simulations on eight multilabel datasets and evaluation on one real-world crowdsourcing dataset consistently show that the proposed OCLDAC significantly outperforms the state-of-the-art methods and their variants.