General Learning Research Articles

DPRA: a dual pooling attention network for point cloud classification and segmentation

Abstract Point cloud data, widely used in fields such as autonomous driving and robotic navigation, involves classification and segmentation tasks. The extraction of local and global features has become a major research focus. The paper proposes two modules: the Dual Pooling Attention model (DP-Attention) and the Residual Attention Module model (RA-MLP). DP-Attention utilizes max pooling and average pooling to compute attention, extracting information between points as well as between features to enhance local feature extraction. RA-MLP integrates self-attention and residual connections to improve global feature extraction. These modules are combined to construct the DPRA network, which is tailored for point cloud classification and segmentation tasks. The DPRA network architecture is based on the encoder-decoder structure of U-Net, using DP-Attention and RA-MLP in the encoder, and only RA-MLP in the decoder. Experimental results on three datasets demonstrate that the DPRA network achieves outstanding performance in the synthetic ModelNet40 classification experiment, with a mean class accuracy of 91.6%. It also achieves the highest mean class accuracy (85.3%) and overall accuracy (86.1%) in the real-world ScanObjectNN classification experiment. Additionally, it attains the highest mean Intersection over Union (mIoU) of 85.2% in the synthetic ShapeNet segmentation experiment. These results indicate that DPRA is applicable to various tasks, demonstrating strong generalization, robustness, and multi-task learning capabilities.

Feature Enhancement-Based Few-Shot Bearing Surface Defect Image Classification Method

The application of intelligent bearing surface defect classification based on deep neural networks remains challenging in real factories, due to the scarcity of defect samples. Under real-working conditions, with less training samples, the paper proposes a few-shot bearing defect image classification network which can recognize different bearing surface defects image, including notch, reddish rust, scratching, incising, conformity, pitting and mill scale. Based on general metric learning neural network framework, a local feature extraction layer is designed, which calculates the auto-correlation vector of global feature in a sliding region to enhance detail features. Additionally, a similar feature attention module emphasizes the the regions of similarity between the query set and the class prototype center to overcome the influence of background noise on classification. To validate the effectiveness of the proposed network, comparative experiments were conducted using the benchmark dataset miniImageNet, achieving classification accuracies of 59% in the 5-way 1-shot setting and 76% in the 5-way 5-shot setting respectively. Furthermore, to assess its performance in a real-factory condition, a self-made dataset of bearing defects from a factory was employed. The proposed network achieved a remarkable classification accuracy of 88% in the 5-way 5-shot setting. These experimental results confirm the practical application value of our few-shot bearing surface defect image classification network, demonstrating its ability to accurately recognize various bearing defects with limited training samples.

Unpaired learning for digital holographic reconstruction and generation.

Traditional numerical reconstruction methods in digital holography (DH) are faced with problems such as inaccurate and time-consuming unwrapping or the need to capture multiple holograms with different diffraction distances. In recent years, deep learning, believed to be a new and effective optimization tool, has been widely used in digital holography. However, most supervised deep learning methods require large-scale paired data, and their preparation is time-consuming and laborious. Here, we propose what we believe to be a new deep learning approach that can use less unpaired data to train neural networks, thereby reducing the need for labeled data. By using CycleGAN, the calculation process of the loss function does not need paired data, and the network can learn the mapping knowledge between two domains, instead of the data and its corresponding label. As such, the proposed method can reconstruct complex amplitudes for holographic reconstruction. And more importantly, the trained network learns the knowledge of the imaging system, leading to the capability to generate experimentally synthetic holograms at the same time. To the best of our knowledge, this is the first demonstration of such high-quality physics-based experimental hologram simulator in DH. The reconstructed complex amplitudes have higher image quality, while the generated holograms can represent realistic system status and reconstruct the input complex amplitudes.

Effects of described demonstrator ability on brain and behavior when learning from others

Observational learning enables us to make decisions by watching others’ behaviors. The quality of such learning depends on the abilities of those we observe, but also on our beliefs about those abilities. We have previously demonstrated that observers learned better from demonstrators described as high vs. low in ability, regardless of their actual performance. The current study aimed to conceptually replicate these findings, and explore the neural mechanisms involved. Forty-five participants performed an observational learning task while undergoing functional magnetic resonance imaging (fMRI). We hypothesized that participants would perform better when demonstrators were described as having high vs. low ability. Unexpectedly, participants performed equally well regardless of described demonstrator ability. The behavioral effects of biased observational learning seem to be driven by mentalizing processes together with general learning and decision-making processes.

Multi-Modal Fusion of Routine Care Electronic Health Records (EHR): A Scoping Review

Background: Electronic health records (EHR) are now widely available in healthcare institutions to document the medical history of patients as they interact with healthcare services. In particular, routine care EHR data are collected for a large number of patients.These data span multiple heterogeneous elements (i.e., demographics, diagnosis, medications, clinical notes, vital signs, and laboratory results) which contain semantic, concept, and temporal information. Recent advances in generative learning techniques were able to leverage the fusion of multiple routine care EHR data elements to enhance clinical decision support. Objective: A scoping review of the proposed techniques including fusion architectures, input data elements, and application areas is needed to synthesize variances and identify research gaps that can promote re-use of these techniques for new clinical outcomes. Design: A comprehensive literature search was conducted using Google Scholar to identify high impact fusion architectures over multi-modal routine care EHR data during the period 2018 to 2023. The guidelines from the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) extension for scoping review were followed. The findings were derived from the selected studies using a thematic and comparative analysis. Results: The scoping review revealed the lack of standard definition for EHR data elements as they are transformed into input modalities. These definitions ignore one or more key characteristics of the data including source, encoding scheme, and concept level. Moreover, in order to adapt to emergent generative learning techniques, the classification of fusion architectures should distinguish fusion from learning and take into consideration that learning can concurrently happen in all three layers of new fusion architectures (i.e., encoding, representation, and decision). These aspects constitute the first step towards a streamlined approach to the design of multi-modal fusion architectures for routine care EHR data. In addition, current pretrained encoding models are inconsistent in their handling of temporal and semantic information thereby hindering their re-use for different applications and clinical settings. Conclusions: Current routine care EHR fusion architectures mostly follow a design-by-example methodology. Guidelines are needed for the design of efficient multi-modal models for a broad range of healthcare applications. In addition to promoting re-use, these guidelines need to outline best practices for combining multiple modalities while leveraging transfer learning and co-learning as well as semantic and temporal encoding.

Anomaly Detection Framework With Contrastive Learning and Multiview Augmentation for Time-Series Domain Generalization

Anomaly Detection Framework With Contrastive Learning and Multiview Augmentation for Time-Series Domain Generalization

Advancements in Clinical Evaluation and Regulatory Frameworks for AI-Driven Software as a Medical Device (SaMD).

Owing to the rapid progress in artificial intelligence (AI) and the widespread use of generative learning, the problem of sparse data has been solved effectively in various research fields. The application of AI technologies has resulted in important transformations in healthcare, particularly in radiology. To ensure the high quality, safety, and effectiveness of AI and machine learning (ML) medical devices, the US Food and Drug Administration (FDA) has established regulatory guidelines to support the performance evaluation of medical devices. Furthermore, the FDA has proposed continuous surveillance requirements for AI/ML medical devices. This paper presents a summary of SaMD products that have passed the FDA 510 (k) AI/ML pathway, the challenges associated with the current AI/ML software-as-a-medical-device, and solutions for promoting the development of AI technologies in medicine. We hope to provide valuable information pertaining to medical-device design, development, and monitoring to ultimately achieve safer and more effective personalized medical services.

MADeGen: Multi-Agent based Deep Reinforcement Learning for Sequential Keyphrase Generation

MADeGen: Multi-Agent based Deep Reinforcement Learning for Sequential Keyphrase Generation

Study on the Role of AI in Transforming HR Practices in Pune's IT Industry

This research study examines the evolving role of AI (Artificial Intelligence) in human resources, aiming on how IT employees perceive and engage with AI Automation at their workplace. It highlights the impact of AI technologies on employee identity, job security, and job satisfaction, while noting significant research gaps in understanding AI's influence on workforce dynamics. The literature classifies AI into categories like general AI tools, machine learning, and robotics, each affecting workplace interactions differently. A major concern is the "dark side" of AI, which may reduce employee autonomy and transparency, impacting their identity and sense of purpose. The research emphasizes AI's transformative potential in HR functions such as recruitment, training & development, performance evaluation and compliance management but also highlights challenges like resistance to change and ethical concerns about AI tools. It advocates for longitudinal studies to better understand AI's long-term effects and suggests that demographic factors may shape employee responses to AI. In conclusion, while AI could enhance HR operations, many questions remain about effective implementation of AI tools in HR functions. Future research should address these gaps to enable organizations to leverage AI for a more innovative and adaptable workforce.

Effect of two teaching methods on nursing students' acquisition of patient-centered communication competence in older people care: a cluster randomized trial.

Patient-centered communication is an essential skill in nursing, particularly in the care of older adult patients. However, generation Z nursing students, who primarily communicate through digital platforms, face unique challenges in adapting to traditional face-to-face communication with older adults. As a result, there is a need for teaching methods that align with this generation's learning style to enhance their communication skills. This study aimed to compare the effectiveness of two teaching methods-standardized patient simulation and role-play-on nursing students' acquisition of patient-centered communication competence in older people care. A controlled cluster-randomized trial was conducted with 124 nursing students, divided into eight teaching groups. Students participated in either a standardized patient simulation or a role-play workshop, each consisting of a 1.5-h online module and a 1.5-h face-to-face session. The three components of patient-centered communication competence-knowledge, skills, and self-efficacy-were assessed using simulated scenarios at pre-test, post-test, and 6-week follow-up. Between-group and within-group differences were measured based on the number of students who achieved competence. Both interventions significantly improved students' knowledge, skills, and self-efficacy in patient-centered communication between pre- and post-tests, with improvements maintained at follow-up. No significant differences were found between the two methods. Both standardized patient simulation and role-play are effective in enhancing patient-centered communication competence in older people care. However, neither method was found to be superior in teaching knowledge, skills, or self-efficacy.

Photonic quantum generative adversarial networks for classical data

In generative learning, models are trained to produce new samples that follow the distribution of the target data. These models were historically difficult to train, until proposals such as generative adversarial networks (GANs) emerged, where a generative and a discriminative model compete against each other in a minimax game. Quantum versions of the algorithm have since been designed for the generation of both classical and quantum data. While most work so far has focused on qubit-based architectures, in this article we present a quantum GAN based on linear optical circuits and Fock-space encoding, which makes it compatible with near-term photonic quantum computing. We demonstrate that the model can learn to generate images by training the model end-to-end experimentally on a single-photon quantum processor.

AUDIO TO SIGN LANGUAGE CONVERSION

Due to high demand for inclusive technologies and the rapid electrons communication age, there is growing interest in automatic audio to sign language conversion systems. This article has given the relevant way to solve this problem by converting spoken language into sign languages in real time using machine translation that can shape communication path for deaf and dumb community. In our method, we use Natural Language Processing (NLP) to transcribe audio into text and then deep learning for sign language generation. The system recognizes keywords and important phrases in the transcribed text, associates them with corresponding sign language gestures, and presents such signs animations using an avatar-based visualization. The recent accomplishment is different from previous systems that work directly and only translate a sentence to its corresponding sign, rather focusing on the contextual analysis for better conveying of meaning. Key Words: Audio to Sign Language Translation, Speech- to- Text Conversion, Natural Language Processing (NLP), Machine Translation, Deep Learning, Gesture Recognition, Avatar-based Sign Language Visualization, Accessibility, Multilingual Sign Language

Closing the educational gap for first generation learners: an exploratory study of interactive learning for underprivileged children in India

Previous research studies have indicated that interactive tools where information is presented through computer mediated technology using elements like text, still images, graphics, animation, video, audio, etc., increase student performance as students (men and women) give more attention and enjoy lessons taught with interactive tools. In India, an increasing number of students belonging to disadvantaged classes withdraw from schools due to a lot of contributing factors such as inability to afford education, disinterest of parents, access to transportation, etc. This study, by applying a mix-methods approach, explores the potential of an interactive learning prototype of a website, in increasing engagement, retention and learning level of grade 4 students who are first generation students. In-class observation and semi-structured interviews, tests, and a questionnaire were used as instruments to collect qualitative and quantitative data from 58 participants. Two types of groups, experimental and control, were examined in the Environmental Science and Maths classes. Experimental groups were taught with the website while control groups were taught in a traditional class. Results show the average test scores of experimental groups in both the subjects to be higher than the control groups. Students preferred learning through interactive tools with colourful images and animations.

DeepPD: A Deep Learning Method for Predicting Peptide Detectability Based on Multi-feature Representation and Information Bottleneck.

Peptide detectability measures the relationship between the protein composition and abundance in the sample and the peptides identified during the analytical procedure. This relationship has significant implications for the fundamental tasks of proteomics. Existing methods primarily rely on a single type of feature representation, which limits their ability to capture the intricate and diverse characteristics of peptides. In response to this limitation, we introduce DeepPD, an innovative deep learning framework incorporating multi-feature representation and the information bottleneck principle (IBP) to predict peptide detectability. DeepPD extracts semantic information from peptides using evolutionary scale modeling 2 (ESM-2) and integrates sequence and evolutionary information to construct the feature space collaboratively. The IBP effectively guides the feature learning process, minimizing redundancy in the feature space. Experimental results across various datasets demonstrate that DeepPD outperforms state-of-the-art methods. Furthermore, we demonstrate that DeepPD exhibits competitive generalization and transfer learning capabilities across diverse datasets and species. In conclusion, DeepPD emerges as the most effective method for predicting peptide detectability, showcasing its potential applicability to other protein sequence prediction tasks.

Backdoor Attacks and Defenses Targeting Multi-Domain AI Models: A Comprehensive Review

Since the emergence of security concerns in artificial intelligence (AI), there has been significant attention devoted to the examination of backdoor attacks. Attackers can utilize backdoor attacks to manipulate model predictions, leading to significant potential harm. However, current research on backdoor attacks and defenses in both theoretical and practical fields still has many shortcomings. To systematically analyze these shortcomings and address the lack of comprehensive reviews, this article presents a comprehensive and systematic summary of both backdoor attacks and defenses targeting multi-domain AI models. Simultaneously, based on the design principles and shared characteristics of triggers in different domains and the implementation stages of backdoor defense, this article proposes a new classification method for backdoor attacks and defenses. We use this method to extensively review backdoor attacks in the fields of computer vision and natural language processing, and we also examine the current applications of backdoor attacks in audio recognition, video action recognition, multimodal tasks, time series tasks, generative learning, and reinforcement learning, while critically analyzing the open problems of various backdoor attack techniques and defense strategies. Finally, this article builds upon the analysis of the current state of AI security to further explore potential future research directions for backdoor attacks and defenses.

MSOTL: Multi-source online transfer learning-enabled performance degradation assessment of machinery

Performance degradation assessment is an important component of predictive maintenance for manufacturing systems and is crucial in determining whether they will persist to the overhaul check stage. However, different manufacturing systems’ feature space varies widely across service and degradation processes. Moreover, in real industrial scenarios, data arrives in real-time in a sequential manner. There is a bottleneck in how the model is being updated in real-time when the data is acquired online. To address the above issues, a novel degradation performance assessment method based on online transfer learning (OTL-DPA) is proposed. The initial transformed matrix is calculated between each source domain and partial target domain to obtain the mapped feature space, based on which the k-nearest neighbor (K-NN) classifier is constructed. When the sequential instances in the remaining target domain arrive, corresponding feature representations are obtained using the above-transformed matrices to reduce the distribution discrepancy. The final assessment results are obtained by integrating the multiple K-NN classifiers on mapped feature spaces. The key of OTL-DPA is to update the transformed matrices and classifier weights each time for a certain amount of samples. Experimental evaluation on fault type diagnosis and fault severity assessment cases indicates that OTL-DPA significantly outperforms non-transfer and general transfer learning methods in the bearing online assessment problems where data were obtained randomly or in a time sequence.

Human-in-the-loop active learning for goal-oriented molecule generation

Machine learning (ML) systems have enabled the modelling of quantitative structure–property relationships (QSPR) and structure-activity relationships (QSAR) using existing experimental data to predict target properties for new molecules. These property predictors hold significant potential in accelerating drug discovery by guiding generative artificial intelligence (AI) agents to explore desired chemical spaces. However, they often struggle to generalize due to the limited scope of the training data. When optimized by generative agents, this limitation can result in the generation of molecules with artificially high predicted probabilities of satisfying target properties, which subsequently fail experimental validation. To address this challenge, we propose an adaptive approach that integrates active learning (AL) and iterative feedback to refine property predictors, thereby improving the outcomes of their optimization by generative AI agents. Our method leverages the Expected Predictive Information Gain (EPIG) criterion to select additional molecules for evaluation by an oracle. This process aims to provide the greatest reduction in predictive uncertainty, enabling more accurate model evaluations of subsequently generated molecules. Recognizing the impracticality of immediate wet-lab or physics-based experiments due to time and logistical constraints, we propose leveraging human experts for their cost-effectiveness and domain knowledge to effectively augment property predictors, bridging gaps in the limited training data. Empirical evaluations through both simulated and real human-in-the-loop experiments demonstrate that our approach refines property predictors to better align with oracle assessments. Additionally, we observe improved accuracy of predicted properties as well as improved drug-likeness among the top-ranking generated molecules.Scientific contributionWe present an adaptable framework that integrates AL and human expertise to refine property predictors for goal-oriented molecule generation. This approach is robust to noise in human feedback and ensures that navigating chemical space with human-refined predictors leverages human insights to identify molecules that not only satisfy predicted property profiles but also score highly on oracle models. Additionally, it prioritizes practical characteristics such as drug-likeness, synthetic accessibility, and a favorable balance between exploring diverse chemical space and exploiting similarity to existing training data.

Deep reinforcement learning for community architectural layout generation

Deep reinforcement learning for community architectural layout generation

Modeling and predicting single-cell multi-gene perturbation responses with scLAMBDA.

Understanding cellular responses to genetic perturbations is essential for understanding gene regulation and phenotype formation. While high-throughput single-cell RNA-sequencing has facilitated detailed profiling of heterogeneous transcriptional responses to perturbations at the single-cell level, there remains a pressing need for computational models that can decode the mechanisms driving these responses and accurately predict outcomes to prioritize target genes for experimental design. Here, we present scLAMBDA, a deep generative learning framework designed to model and predict single-cell transcriptional responses to genetic perturbations, including single-gene and combinatorial multi-gene perturbations. By leveraging gene embeddings derived from large language models, scLAMBDA effectively integrates prior biological knowledge and disentangles basal cell states from perturbation-specific salient representations. Through comprehensive evaluations on multiple single-cell CRISPR Perturb-seq datasets, scLAMBDA consistently outperformed state-of-the-art methods in predicting perturbation outcomes, achieving higher prediction accuracy. Notably, scLAMBDA demonstrated robust generalization to unseen target genes and perturbations, and its predictions captured both average expression changes and the heterogeneity of single-cell responses. Furthermore, its predictions enable diverse downstream analyses, including the identification of differentially expressed genes and the exploration of genetic interactions, demonstrating its utility and versatility.

The cerebellum is involved in implicit motor sequence learning.

Implicit motor sequence learning (IMSL) is a cognitive function that allows us to execute multiple movements in a specific sequential order and plays a crucial role in our daily functional activities. Although the role of the basal ganglia network in IMSL is well-established, the exact involvement of the cerebellar network is less clear. Here, we aimed to address this issue by investigating the effects of cerebellar transcranial direct-current stimulation (tDCS) on IMSL. In this sham-controlled, crossover study in 45 healthy young adults, we used mixed-effects models to analyze sequence-specific (primary outcome) and general learning effects (secondary outcome) in the acquisition (during tDCS), short- (five minutes post-tDCS) and long-term consolidation (one week post-tDCS) phases of IMSL, as measured by the serial reaction time (SRT) task. Analyses based on response times (RTs) revealed that anodal tDCS over the cerebellum significantly increased sequence-specific learning during acquisition, compared to sham (anodal: M = 38.24 ms, sham: M = 26.78 ms, p = 0.032); did not affect general learning; and significantly slowed overall RTs (anodal: M = 362.03 ms, sham: M = 356.37 ms, p = 0.049). Accuracy-based analyses revealed that anodal tDCS reduced the probability of correct responses occurring in random trials versus sequential trials by 1.17%, p = 0.009, whereas sham tDCS had no effect, p = 0.999. Our finding of enhanced sequence-specific learning, but not general learning, suggests that the cerebellar network not only plays a role in error correction processes, but also serves a sequence-specific function within the integrated motor learning network that connects the basal ganglia and cerebellum.