Stable Learning Research Articles

Hybrid-DC: A Hybrid Framework Using ResNet-50 and Vision Transformer for Steel Surface Defect Classification in the Rolling Process

This study introduces Hybrid-DC, a hybrid deep-learning model integrating ResNet-50 and Vision Transformer (ViT) for high-accuracy steel surface defect classification. Hybrid-DC leverages ResNet-50 for efficient feature extraction at both low and high levels and utilizes ViT’s global context learning to enhance classification precision. A unique hybrid attention layer and an attention fusion mechanism enable Hybrid-DC to adapt to the complex, variable patterns typical of steel surface defects. Experimental evaluations demonstrate that Hybrid-DC achieves substantial accuracy improvements and significantly reduced loss compared to traditional models like MobileNetV2 and ResNet, with a validation accuracy reaching 0.9944. The results suggest that this model, characterized by rapid convergence and stable learning, can be applied for real-time quality control in steel manufacturing and other high-precision industries, enhancing automated defect detection efficiency.

JFDI: Joint Feature Differentiation and Interaction for domain adaptive object detection

In unsupervised domain adaptive object detection, learning target-specific features is pivotal in enhancing detector performance. However, previous methods mostly concentrated on aligning domain-invariant features across domains and neglected integrating the specific features. To tackle this issue, we introduce a novel feature learning method called Joint Feature Differentiation and Interaction (JFDI), which significantly boosts the adaptability of the object detector. We construct a dual-path architecture based on we proposed feature differentiate modules: One path, guided by the source domain data, utilizes multiple discriminators to confuse and align domain-invariant features. The other path, specifically tailored to the target domain, learns its distinctive characteristics based on pseudo-labeled target data. Subsequently, we implement an interactive enhanced mechanism between these paths to ensure stable learning of features and mitigate interference from pseudo-label noise during the iterative optimization. Additionally, we devise a hierarchical pseudo-label fusion module that consolidates more comprehensive and reliable results. In addition, we analyze the generalization error bound of JFDI, which provides a theoretical basis for the effectiveness of JFDI. Extensive empirical evaluations across diverse benchmark scenarios demonstrate that our method is advanced and efficient.

Provably Stable Learning Control of Linear Dynamics With Multiplicative Noise

Provably Stable Learning Control of Linear Dynamics With Multiplicative Noise

Discrete Space Deep Reinforcement Learning Algorithm Based on Support Vector Machine Recursive Feature Elimination

Algorithms for training agents with experience replay have advanced in several domains, primarily because prioritized experience replay (PER) developed from the double deep Q-network (DDQN) in deep reinforcement learning (DRL) has become a standard. PER-based algorithms have achieved significant success in the image and video domains. However, the exceptional results observed in images and videos are not as effective in many domains with simple action spaces and relatively small states, particularly in discrete action spaces with sparse rewards. Moreover, most advanced techniques may improve sampling efficiency using deep learning algorithms rather than reinforcement learning. However, there is growing evidence that deep learning algorithms cannot generalize during training. Therefore, this study proposes an algorithm suitable for discrete action space environments that uses the sample efficiency of PER based on DDQN but incorporates support vector machine recursive feature elimination (SVM-RFE) without enhancing the sampling efficiency through deep learning algorithms. The proposed algorithm exhibited considerable performance improvements in classical OpenAI Gym environments that did not use images or videos as inputs. In particular, simple discrete space environments with reflection symmetry, such as Cart–Pole, exhibited a faster and more stable learning process. These results suggest that the application of SVM-RFE, which leverages the orthogonality of support vector machines (SVMs) across learning patterns, can be appropriate when the data in the reinforcement learning environment demonstrate symmetry.

SL: Stable Learning in Source-Free Domain Adaptation for Medical Image Segmentation

Deep learning techniques for medical image analysis often encounter domain shifts between source and target data. Most existing approaches focus on unsupervised domain adaptation (UDA). However, in practical applications, many source domain data are often inaccessible due to issues such as privacy concerns. For instance, data from different hospitals exhibit domain shifts due to equipment discrepancies, and data from both domains cannot be accessed simultaneously because of privacy issues. This challenge, known as source-free UDA, limits the effectiveness of previous UDA medical methods. Despite the introduction of various medical source-free unsupervised domain adaptation (MSFUDA) methods, they tend to suffer from an over-fitting problem described as “longer training, worse performance”. To address this issue, we proposed the Stable Learning (SL) strategy. SL is a method that can be integrated with other approaches and consists of weight consolidation and entropy increase. Weight consolidation helps retain domain-invariant knowledge, while entropy increase prevents over-learning. We validated our strategy through experiments on three MSFUDA methods and two public datasets. For the abdominal dataset, the application of the SL strategy enables the MSFUDA method to effectively address the domain shift issue. This results in an improvement in the Dice coefficient from 0.5167 to 0.7006 for the adaptation from CT to MRI, and from 0.6474 to 0.7188 for the adaptation from MRI to CT. The same improvement is observed with the cardiac dataset. Additionally, we conducted ablation studies on the two involved modules, and the results demonstrated the effectiveness of the SL strategy.

Boundary integrated neural networks for 2D elastostatic and piezoelectric problems

In this paper, we make the first attempt to adopt the boundary integrated neural networks (BINNs) for the numerical solution of two-dimensional (2D) elastostatic and piezoelectric problems. The proposed BINNs combine the artificial neural networks with the exact boundary integral equations (BIEs) to effectively solve the boundary value problems based on the corresponding partial differential equations (PDEs). The BIEs are utilized to localize all the unknown physical quantities on the boundary, which are approximated by using artificial neural networks and resolved via a training process. In contrast to many traditional neural network methods based on a domain discretization, the present BINNs offer several distinct advantages. Firstly, by embedding the analytical BIEs into the learning procedure, the present BINNs only need to discretize the boundary of the problem domain, which reduces the number of the unknowns and can lead to a faster and more stable learning process. Secondly, the differential operators in the original PDEs are substituted by integral operators, which can effectively eliminate the need for additional differentiations of the neural networks (high-order derivatives of the neural networks may lead to instabilities in the learning process). Thirdly, the loss function of the present BINNs only contains the residuals of the BIEs, as all the boundary conditions have been inherently incorporated within the formulation. Therefore, there is no necessity for employing any weighting functions, which are commonly used in most traditional methods to balance the gradients among the different objective functions. Extensive numerical experiments show that the present BINNs are much easier to train and can usually provide more accurate solutions as compared to many traditional neural network methods.

UDC 376.07:005.95(477-072)“364” FEATURES OF MANAGEMENT OF SPECIAL AND INCLUSIVE EDUCATION INSTITUTIONS IN THE FRONTLINE REGION

The article examines the current challenges faced by the heads of special and inclusive educational institutions in the context of ongoing hostilities in Ukraine. The aim of the article is to analyze and highlight these challenges, with an emphasis on the need to modernize management approaches in response to the growing influence of inclusive processes. The authors emphasize the critical importance of forming a new generation of educational leaders who have the necessary competencies to manage innovative educational institutions that implement inclusive approaches. This includes the recognition of universal values, ensuring the humane treatment of people with disabilities, as well as addressing the issues of their socialization and adaptation in society as a whole. Strategic directions of inclusive education require leaders who are not only aware, but also actively engaged in creating and maintaining environments where both healthy pupils and students with special educational needs can study together in specialized institutions. The article emphasizes the need for these leaders to be well aware of the latest trends in the development of special and inclusive education and to effectively manage the complex dynamics that arise in the frontline regions. In these regions, the challenges are particularly acute, with security, infrastructure preservation, staffing and adequate funding coming to the fore. The constant threat to life due to hostilities poses significant risks to the physical and psychological well-being of students and staff. The destruction of infrastructure, including educational facilities, exacerbates the difficulties of maintaining a stable learning environment. Communication problems, whether due to damaged networks or broken communication lines, make it even more difficult to manage these institutions. Staff shortages, often caused by the relocation of staff or the inability to attract qualified specialists to dangerous areas, add another layer of complexity to the management task. The article also highlights the problem of professional burnout, which is exacerbated by stressful working conditions in the conflict zone. In addition, frequent interruptions in the educational process, whether due to evacuations, damage to facilities, or other war-related disruptions, significantly complicate the ability of these institutions to fulfill their educational mission. The article concludes by emphasizing the need for rapid adaptation and innovative solutions on the part of education managers to effectively address these challenges, ensuring that inclusive education can continue to progress even in the most adverse conditions.

Research on VR e-learning based on load balancing algorithm in intelligent music education system

This paper aims to explore the application of virtual reality (VR) electronic learning based on load balancing algorithm in the simulation of intelligent music education system. This paper designs and implements an intelligent music education system simulation, which uses load balancing algorithm and virtual reality (VR) technology to provide the convenience of e-learning and personalized learning experience. The load balancing algorithm allocates and manages multiple tasks or requests in the system to ensure the balanced utilization of resources and the efficient operation of the system. In the intelligent music education system, the load balancing algorithm can be used to manage user requests, allocate appropriate resources and services, and ensure that the system can respond quickly and provide a stable learning environment. This paper introduces the architecture and implementation details of the system, including the design and optimization of load balancing algorithm and the application of VR technology. The results show that we point out that VR e-learning based on load balancing algorithm has potential in the simulation of intelligent music education system, which can provide high-quality personalized music learning experience and provide useful enlightenment for innovation in the field of music education.

Enhancing stability and interpretability in the study of strength behavior for coarse-grained soils

Coarse-grained soils have been commonly employed as filling materials in engineering projects, and comprehending the mechanical properties of these soils is crucial for ensuring project safety. Currently, artificial intelligence (AI) plays a significant role in advancing these fields. However, the inherent lack of stability and interpretability of AI models constrains their potential for providing substantial assistance. To solve these shortcomings, this study investigates the mechanical properties of coarse-grained soils and proposes an approach that integrates a stable learning module with neural networks to enhance the stability of an AI model. The proposed method addresses the challenge of relaxing the strong assumption of independent and identically distributed (i.i.d.) data. In addition, it improves the stability of the prediction of a relationship between the deviatoric stress and axial strain. The proposed method is verified by experiments. The experimental results confirm that the proposed method performs well on real-world data samples that deviate from the i.i.d. assumption. Moreover, the proposed method enhances the interpretability of the AI model and facilitates the identification of crucial features for assessing the strength behavior of coarse-grained soils. This feature of interpretability is invaluable for practical engineering applications.

Humans adaptively select different computational strategies in different learning environments.

The Rescorla-Wagner rule remains the most popular tool to describe human behavior in reinforcement learning tasks. Nevertheless, it cannot fit human learning in complex environments. Previous work proposed several hierarchical extensions of this learning rule. However, it remains unclear when a flat (nonhierarchical) versus a hierarchical strategy is adaptive, or when it is implemented by humans. To address this question, current work applies a nested modeling approach to evaluate multiple models in multiple reinforcement learning environments both computationally (which approach performs best) and empirically (which approach fits human data best). We consider 10 empirical data sets (N = 407) divided over three reinforcement learning environments. Our results demonstrate that different environments are best solved with different learning strategies; and that humans adaptively select the learning strategy that allows best performance. Specifically, while flat learning fitted best in less complex stable learning environments, humans employed more hierarchically complex models in more complex environments. (PsycInfo Database Record (c) 2024 APA, all rights reserved).

Generating lymphoma ultrasound image description with transformer model

Lymphoma, the most prevalent hematologic tumor originating from the lymphatic hematopoietic system, can be accurately diagnosed using high-resolution ultrasound. Microscopic ultrasound performance enables clinicians to identify suspected tumors and subsequently obtain a definitive pathological diagnosis through puncture biopsy. However, the complex and diverse ultrasonographic manifestations of lymphoma pose challenges for accurate characterization by sonographers. To address these issues, this study proposes a Transformer-based model for generating descriptive ultrasound images of lymphoma, aiming to provide auxiliary guidance for ultrasound doctors during screening procedures. Specifically, deep stable learning is integrated into the model to eliminate feature dependencies by training sample weights. Additionally, a memory module is incorporated into the model decoder to enhance semantic information modeling in descriptions and utilize learned semantic tree branch structures for more detailed image depiction. Experimental results on an ultrasonic diagnosis dataset from Shanghai Ruijin Hospital demonstrate that our proposed model outperforms relevant methods in terms of prediction performance.

Actively learning costly reward functions for reinforcement learning

Transfer of recent advances in deep reinforcement learning to real-world applications is hindered by high data demands and thus low efficiency and scalability. Through independent improvements of components such as replay buffers or more stable learning algorithms, and through massively distributed systems, training time could be reduced from several days to several hours for standard benchmark tasks. However, while rewards in simulated environments are well-defined and easy to compute, reward evaluation becomes the bottleneck in many real-world environments, e.g. in molecular optimization tasks, where computationally demanding simulations or even experiments are required to evaluate states and to quantify rewards. When ground-truth evaluations become orders of magnitude more expensive than in research scenarios, direct transfer of recent advances would require massive amounts of scale, just for evaluating rewards rather than training the models. We propose to alleviate this problem by replacing costly ground-truth rewards with rewards modeled by neural networks, counteracting non-stationarity of state and reward distributions during training with an active learning component. We demonstrate that using our proposed method, it is possible to train agents in complex real-world environments orders of magnitudes faster than would be possible when using ground-truth rewards. By enabling the application of RL methods to new domains, we show that we can find interesting and non-trivial solutions to real-world optimization problems in chemistry, materials science and engineering. We demonstrate speed-up factors of 50–3000 when applying our approach to challenges of molecular design and airfoil optimization.

BCM‐Inspired Synapses Constructed with Barrier‐Modulated Coupling Junctions for Enhancing Speech Recognition

AbstractBio‐inspired synaptic devices have garnered considerable interest in neuromorphic computing. The Bienenstock‐Cooper‐Munro (BCM) learning rule stands out as one of the most accurate synaptic models, featuring non‐monotonic behavior and threshold sliding effect, crucial for stable learning processes. The direct device strategy for completely mimicking the BCM rule is a tough issue since the current devices lack two competitive working modes within one device. In this work, a dual‐junction synaptic device with opposite built‐in electric fields using a W/WO2/WO3‐x/Au structure is demonstrated. The devices directly mimic two fundamental features of the BCM rule via a delicately‐designed bandgap engineering strategy. Furthermore, the working mechanisms are investigated and the promising potential of dual‐junction synaptic devices is demonstrated for enhancing speech recognition through Convolutional Neural Network (CNN)‐based digital speech recognition with a remarkable accuracy of 98% through a synaptic array. Even for speech recognition with 13% Gaussian noise, the accuracy remained at 83%. These findings provide a promising strategy for developing BCM‐based synaptic devices for neuromorphic computing applications.

Machine learning techniques for diagrid building design: Architectural–Structural correlations with feature selection and data augmentation

Artificial intelligence (AI) and machine learning (ML) techniques are transforming building engineering. This work goes through the critical role of architectural parameters in influencing the structural responses of tall buildings, with a special focus on diagrid structures. The main aim of this study is to demonstrate how ML can improve the early design phase of diagrid buildings. Using a small, initially collected data set, enhanced through data augmentation, the classification of diagrid buildings in terms of design feasibility is investigated. This study identifies key architectural and structural parameters, employing various filter and wrapper methods for feature selection. The results show that our methods are effective in producing high-quality synthetic data, maintaining stable learning accuracies, and establishing accurate and robust relationships between architectural parameters and structural responses in diagrid buildings. These insights are crucial for facilitating more effective design processes in the realm of high-rise diagrid building design.

State-Dependent Maximum Entropy Reinforcement Learning for Robot Long-Horizon Task Learning

Task-oriented robot learning has shown significant potential with the development of Reinforcement Learning (RL) algorithms. However, the learning of long-horizon tasks for robots remains a formidable challenge due to the inherent complexity of tasks, typically comprising multiple diverse stages. Universal RL algorithms commonly encounter issues such as slow convergence or even failure to converge altogether when applied to such tasks. The reasons behind these challenges lie in the local optima trap or redundant exploration during the new stages or the junction of two continuous stages. To address these challenges, we propose a novel state-dependent maximum entropy (SDME) reinforcement learning algorithm. This algorithm effectively balances the trade-off between exploration and exploitation around three kinds of critical states arising from the unique nature of long-horizon tasks. We conducted experiments within an open-source simulation environment, focusing on two representative long-horizon tasks. The proposed SDME algorithm exhibits faster and more stable learning capabilities, requiring merely one-third of the number of learning samples necessary for baseline approaches. Furthermore, we assess the generalization ability of our method under randomly initialized conditions, and the results show that the success rate of the SDME algorithm is nearly twice that of the baselines. Our code will be available at https://github.com/Peter-zds/SDME.

State observer-based Physics-Informed Machine Learning for leader-following tracking control of mobile robot

In this paper, the novel leader-following tracking control method is proposed for mobile robots, which consists estimation technique of the speed of the leader robot (LR), and a parameter-dependent controller for the follower robot (FR). To estimate the speed of LR, a novel Physics Informed Machine Learning (PIML) is proposed to learn the dynamics of the state observer via the error state model. The dynamics of the state observer in PIML play a significant role for stable learning and state estimation of uncertain models. The gain of the parameter-dependent controller is determined by the convex combination of the robust control technique via the polytopic model. Finally, the tracking performance of the proposed method is verified through the simulation and experiment.

An Intelligent Speech Recognition Method Based on Stable Learning

An Intelligent Speech Recognition Method Based on Stable Learning

Stable Learning via Dual Feature Learning

Stable Learning via Dual Feature Learning

Stable Learning via Triplex Learning

Stable Learning via Triplex Learning

Software Defect Prediction Method Based on Stable Learning

Software Defect Prediction Method Based on Stable Learning