A comparison review of transfer learning and self-supervised learning: Definitions, applications, advantages and limitations

Expert Systems with Applications, Год журнала: 2023, Номер 242, С. 122807 - 122807

Опубликована: Дек. 2, 2023

Deep learning has emerged as a powerful tool in various domains, revolutionising machine research. However, one persistent challenge is the scarcity of labelled training data, which hampers performance and generalisation deep models. To address this limitation, researchers have developed innovative methods to overcome data enhance model capabilities. Two prevalent techniques that gained significant attention are transfer self-supervised learning. Transfer leverages knowledge learned from pre-training on large-scale dataset, such ImageNet, applies it target task with limited data. This approach allows models benefit representations effectively new tasks, resulting improved generalisation. On other hand, focuses using pretext tasks do not require manual annotation, allowing them learn valuable large amounts unlabelled These can then be fine-tuned for downstream mitigating need extensive In recent years, found applications fields, including medical image processing, video recognition, natural language processing. approaches demonstrated remarkable achievements, enabling breakthroughs areas disease diagnosis, object understanding. while these offer numerous advantages, they also limitations. For example, may face domain mismatch issues between requires careful design ensure meaningful representations. review paper explores fields within past three years. It delves into advantages limitations each approach, assesses employing techniques, identifies potential directions future By providing comprehensive current methods, article offers guidance selecting best technique specific issue.

Язык: Английский

---

Artificial intelligence and multimodal data fusion for smart healthcare: topic modeling and bibliometrics

Artificial Intelligence Review, Год журнала: 2024, Номер 57(4)

Опубликована: Март 15, 2024

Abstract Advancements in artificial intelligence (AI) have driven extensive research into developing diverse multimodal data analysis approaches for smart healthcare. There is a scarcity of large-scale literature this field based on quantitative approaches. This study performed bibliometric and topic modeling examination 683 articles from 2002 to 2022, focusing topics trends, journals, countries/regions, institutions, authors, scientific collaborations. Results showed that, firstly, the number has grown 1 220 with majority being published interdisciplinary journals that link healthcare medical information technology AI. Secondly, significant rise quantity can be attributed increasing contribution scholars non-English speaking countries/regions noteworthy contributions made by authors USA India. Thirdly, researchers show high interest issues, especially, cross-modality magnetic resonance imaging (MRI) brain tumor analysis, cancer prognosis through multi-dimensional AI-assisted diagnostics personalization healthcare, each experiencing increase interest. an emerging trend towards issues such as applying generative adversarial networks contrastive learning image fusion synthesis utilizing combined spatiotemporal resolution functional MRI electroencephalogram data-centric manner. valuable enhancing researchers’ practitioners’ understanding present focal points upcoming trajectories AI-powered analysis.

Язык: Английский

---

Trustworthy deep learning framework for the detection of abnormalities in X-ray shoulder images

PLoS ONE, Год журнала: 2024, Номер 19(3), С. e0299545 - e0299545

Опубликована: Март 11, 2024

Musculoskeletal conditions affect an estimated 1.7 billion people worldwide, causing intense pain and disability. These lead to 30 million emergency room visits yearly, the numbers are only increasing. However, diagnosing musculoskeletal issues can be challenging, especially in emergencies where quick decisions necessary. Deep learning (DL) has shown promise various medical applications. previous methods had poor performance a lack of transparency detecting shoulder abnormalities on X-ray images due training data better representation features. This often resulted overfitting, generalisation, potential bias decision-making. To address these issues, new trustworthy DL framework been proposed detect (such as fractures, deformities, arthritis) using images. The consists two parts: same-domain transfer (TL) mitigate imageNet mismatch feature fusion reduce error rates improve trust final result. Same-domain TL involves pre-trained models large number labelled from body parts fine-tuning them target dataset Feature combines extracted features with seven train several ML classifiers. achieved excellent accuracy rate 99.2%, F1 Score Cohen’s kappa 98.5%. Furthermore, results was validated three visualisation tools, including gradient-based class activation heat map (Grad CAM), visualisation, locally interpretable model-independent explanations (LIME). outperformed orthopaedic surgeons invited classify test set, who obtained average 79.1%. proven effective robust, improving generalisation increasing results.

Язык: Английский

---

ATD Learning: A secure, smart, and decentralised learning method for big data environments

Information Fusion, Год журнала: 2025, Номер unknown, С. 102953 - 102953

Опубликована: Янв. 1, 2025

Язык: Английский

---

Towards Risk-Free Trustworthy Artificial Intelligence: Significance and Requirements

International Journal of Intelligent Systems, Год журнала: 2023, Номер 2023, С. 1 - 41

Опубликована: Окт. 26, 2023

Given the tremendous potential and influence of artificial intelligence (AI) algorithmic decision-making (DM), these systems have found wide-ranging applications across diverse fields, including education, business, healthcare industries, government, justice sectors. While AI DM offer significant benefits, they also carry risk unfavourable outcomes for users society. As a result, ensuring safety, reliability, trustworthiness becomes crucial. This article aims to provide comprehensive review synergy between DM, focussing on importance trustworthiness. The addresses following four key questions, guiding readers towards deeper understanding this topic: (i) why do we need trustworthy AI? (ii) what are requirements In line with second question, that establish been explained, explainability, accountability, robustness, fairness, acceptance AI, privacy, accuracy, reproducibility, human agency, oversight. (iii) how can data? (iv) priorities in terms challenging applications? Regarding last six different discussed, environmental science, 5G-based IoT networks, robotics architecture, engineering construction, financial technology, healthcare. emphasises address before their deployment order achieve goal good. An example is provided demonstrates be employed eliminate bias resources management systems. insights recommendations presented paper will serve as valuable guide researchers seeking applications.

Язык: Английский

---

MEFF – A model ensemble feature fusion approach for tackling adversarial attacks in medical imaging

Intelligent Systems with Applications, Год журнала: 2024, Номер 22, С. 200355 - 200355

Опубликована: Март 16, 2024

Adversarial attacks pose a significant threat to deep learning models, specifically medical images, as they can mislead models into making inaccurate predictions by introducing subtle distortions the input data that are often imperceptible humans. Although adversarial training is common technique used mitigate these on it lacks flexibility address new attack methods and effectively improve feature representation. This paper introduces novel Model Ensemble Feature Fusion (MEFF) designed combat in image applications. The proposed model employs fusion combining features extracted from different DL then trains Machine Learning classifiers using fused features. It uses concatenation method merge features, forming more comprehensive representation enhancing model's ability classify classes accurately. Our experimental study has performed evaluation of MEFF, considering several challenging scenarios, including 2D 3D greyscale colour binary classification, multi-label classification. reported results demonstrate robustness MEFF against types across six distinct A key advantage its capability incorporate wide range without need train scratch. Therefore, contributes developing diverse robust defense strategy. More importantly, leveraging ensemble modeling, enhances resilience face attacks, paving way for improved reliability analysis.

Язык: Английский

---

SSP: self-supervised pertaining technique for classification of shoulder implants in x-ray medical images: a broad experimental study

Artificial Intelligence Review, Год журнала: 2024, Номер 57(10)

Опубликована: Авг. 18, 2024

Abstract Multiple pathologic conditions can lead to a diseased and symptomatic glenohumeral joint for which total shoulder arthroplasty (TSA) replacement may be indicated. The long-term survival of implants is limited. With the increasing incidence surgery, it anticipated that revision surgery will become more common. It challenging at times retrieve manufacturer in situ implant. Therefore, certain systems facilitated by AI techniques such as deep learning (DL) help correctly identify implanted prosthesis. Correct identification reduce perioperative complications complications. DL was used this study categorise different based on X-ray images into four classes (as first case small dataset): Cofield, Depuy, Tornier, Zimmer. Imbalanced public datasets poor performance model training. Most methods literature have adopted idea transfer (TL) from ImageNet models. This type TL has been proven ineffective due some concerns regarding contrast between features learnt natural (ImageNet: colour images) (greyscale images). To address that, new approach (self-supervised pertaining (SSP)) proposed resolve issue datasets. SSP training models (ImageNet models) large number unlabelled greyscale medical domain update features. are then trained labelled data set implants. shows excellent results five models, including MobilNetV2, DarkNet19, Xception, InceptionResNetV2, EfficientNet with precision 96.69%, 95.45%, 98.76%, 98.35%, 96.6%, respectively. Furthermore, shown domains (such ImageNet) do not significantly affect images. A lightweight scratch achieves 96.6% accuracy, similar using standard extracted train several ML classifiers show outstanding obtaining an accuracy 99.20% Xception+SVM. Finally, extended experimentation carried out elucidate our approach’s real effectiveness dealing imaging scenarios. Specifically, tested without SSP, 99.47% CT brain stroke 98.60%.

Язык: Английский

---

Real Time Abnormal Behavior Detection and Warning System Based on Deep Convolutional Neural Network

SSRN Electronic Journal, Год журнала: 2025, Номер unknown

Опубликована: Янв. 1, 2025

With offensive activities on the rise we have never needed safety measures more. Closed-circuit television (CCTV) is now being deployed regularly in public spaces these days where better need to be taken such as shopping malls, banks and other high-traffic places. But manually watching cameras too complicated one must keep their eyes it every single minute all suspicious are hard catch. In addressing this problem, present research offers a fully functional system popularly referred Suspicious Activity Recognition System (SARS), which employs state-of-the-art deep learning technologies. It tries automatically monitor real-time violent clues from video feeds by removing out irritated behaviours. detects high movement find using different models. Most definitely, for protecting humans road soon incident takes place, alerts will sent through time something unusual. This method actually tightens security at same reduces human operator effort.

Язык: Английский

---

Employing the Concept of Stacking Ensemble Learning to Generate Deep Dream Images Using Multiple CNN Variants

Intelligent Systems with Applications, Год журнала: 2025, Номер unknown, С. 200488 - 200488

Опубликована: Янв. 1, 2025

Язык: Английский

---

Multi-category sensitive image recognition based on RefCA-EfficientNetV2

International Journal of General Systems, Год журнала: 2025, Номер unknown, С. 1 - 25

Опубликована: Апрель 10, 2025

Язык: Английский

---