Applications of Generative AI

This chapter Abrahamaims to bridge the gap between the theoretical potential of Generative AI (GAI) tools, such as Generative Pretrained Transformer (GPT), and their practical applications as supportive tools for scientific research. The chapter provides approaches and techniques for leveraging GAI to address research challenges and activities. It describes common research tasks and provides guidance on how to use GPT to solve them. To the best of our experience, at the current stage, the integration of researchers and GPT has the potential to yield better results than either could achieve alone. Furthermore, the increasing availability of GPT tools suggests that the synergy between the two will continue to improve research output quality and save time. The key to successful integration lies in the appropriate use of GPT that can be achieved by directing AI tools to solve tasks effectively and using prompt engineering techniques.

Search campaigns consist of ad groups. An ad group contains a related set of keywords and ads. During an online campaign, search advertisers experiment with different marketing messages such as subtle vs. strong being used in ad copies, with different keywords from broad match to exact match while targeting online users, and with different landing pages to target information seekers vs. product buyers. Generating new ads and keywords for this experimental endeavor becomes essential in effective search campaign management. In this work, we discuss the role of generative AI in creating and managing search ad campaigns programmatically.

Deep generative models have been widely employed across diverse fields, ranging from image and video analysis to natural language processing. In combination with the increasing computational power and abundant data resources available in the public domain, generative models have made significant advancements into the area of drug discovery and development. In particular, generative models are being extensively explored for de novo design of novel molecules, endowed with desirable physicochemical properties or biological activity, thereby accelerating the hit discovery phase by more rapidly sampling the chemical space of drug-like compounds. However, despite their considerable potential, these methods do have limitations that warrant consideration. For instance, they tend to generate compounds that may exhibit chemical instability, pose challenges in synthesis, or bear resemblance to existing drugs, thereby raising concerns regarding patentability. Furthermore, the experimental validation of the generated molecules through exemplary case studies remains limited. This chapter focuses on the application of generative models in de novo drug design. Firstly, we provide a brief introduction to commonly used generative models, such as recurrent neural networks, autoencoders, generative adversarial networks, as well as transfer learning and reinforcement learning techniques. Secondly, we conduct a comprehensive review of the latest developments in utilizing various generative models for drug discovery. This includes an analysis of benchmarks, metrics, and performance evaluation methods through the examination of diverse case studies. Finally, we shed light on the challenges associated with generative methods and discuss future directions in this dynamic and rapidly evolving field.

The high ability of generative models to generate synthetic samples with distribution similar to real data samples brings many benefits in various applications. However, one of the most major elements in the success of generative models is the data that is used to train these models, and preserving privacy of this data is necessary. However, various studies have shown that the high capacity of generative models leads to memorizing the details of the training data by these models, and different attacks have been conducted against generative models which infer information about training data from trained model. Also, many privacy-preserving mechanisms have been proposed to defend against these attacks. In this chapter, after introducing the topic, the privacy attacks against generative models and relevant defense mechanisms are discussed. In particular, the privacy attacks and related privacy preserving methods are categorized and discussed. Then, some challenges and future research directions are examined.

Recently, generative adversarial networks (GANs) have been investigated since 2014, and many algorithmic solutions have been suggested for them. GAN have recently become a popular research topic. Even so, not many studies are deep enough to explain the relationship between the many variations of GAN and how they arise. We aim to provide a survey of different GAN techniques in this work, discussing them from the angles of theory, algorithms, and practical applications. We begin with a comprehensive introduction, architecture, and applications of the most popular GAN algorithms, then we draw parallels and draw distinctions between them. The second part of the study focuses on examining the theoretical issues associated with GANs. In this work, we try to determine the benefits and drawbacks of GANs, as well as the important obstacles that stand in the way of achieving successful implementation of GAN in a variety of application domains. Typical GAN applications are presented, including those in synthetic image processing and computer vision, natural language processing, music, speech, and audio, medicine, and data science. The final section of the research presents the study’s conclusion and some suggestions for further research. Further, highlighting the pros and cons of ongoing studies on the application of aversive learning can help guide future research efforts toward the most fruitful avenues.

This chapter delves into the concepts of neural rendering and generative models, highlighting their importance in the fields of computer graphics, computer vision, and artificial intelligence. Neural rendering techniques utilize deep learning algorithms to generate realistic images, videos, or 3D models, while generative models learn the underlying data distribution to create novel content. The chapter explores various methodologies, such as Neural Radiance Fields, Variational Autoencoders, and Generative Adversarial Networks, among others. Applications of these techniques, such as photorealistic rendering, style transfer, and image synthesis, are discussed, along with the challenges and limitations associated with them. The chapter concludes with an outlook on the future prospects of neural rendering and generative models, emphasizing their potential to revolutionize digital content creation and consumption.

In this chapter, a computational approach is proposed to address the challenge of degraded visibility caused by smoke during laparoscopic surgery. The visualization of organs and tissues is hindered by the presence of smoke, which results from dissection tools. This, in turn, leads to potential errors and increased surgical duration, ultimately impacting patient outcomes. To overcome this issue, a novel neural architecture is introduced, which consists of two autoencoders trained using the generative neural network paradigm. The image segmentation on the laparoscopic image is performed by the first autoencoder, while the second autoencoder incorporates this segmented image as an additional fifth channel. To evaluate the effectiveness of the approach, comprehensive quantitative assessments are conducted, and the results are compared with state-of-the-art desmoking and dehazing techniques. Performance evaluation is carried out using commonly used metrics in the field. The superiority of the proposed method over existing approaches is demonstrated by the obtained results. This makes the method highly suitable for integration into medical systems using embedded devices.

This chapter presents a comprehensive exploration of Generative Artificial Intelligence (AI) and its applications in the Architecture, Engineering, Construction, and Facility Management (AEC-FM) industry. It begins with a general overview of Generative AI, highlighting its capacity to create novel content. The subsequent section delves into the technological requirements for implementing Generative AI in the AEC-FM industry, including Internet of Things (IoT), Distributed Ledger Technology, Computing, Deep Learning, Natural Language Processing (NLP), Knowledge Graph, Computer Vision, and Immersive Technologies. The chapter then proceeds to discuss the wide-ranging applications of Generative AI in the AEC-FM industry, showcasing its potential in enhancing design processes, predictive modeling, operations, and maintenance. Furthermore, the role of Building Information Modeling (BIM) as a facilitator for Generative AI in the AEC-FM industry is explored. Identifying gaps and trends in Generative AI research, the chapter presents a keyword co-occurrence analysis specific to the AEC-FM industry, covering lifecycle management, predictive modeling, machine learning implementation in design, design optimization, constraint-based generative design, and design techniques. Finally, the chapter concludes by examining the future prospects for Generative AI in AEC-FM activities, offering insights into its potential transformative impact on the industry.

The global AI in Healthcare market is expected to grow from USD14.6 billion in 2023 to 102.7 billion by 2028 with a compound annual growth rate over that period of 47.6%. There is universal agreement that there is a rising demand for AI services as population ageing throughout the developed world causes the number of patients to increase faster than the healthcare workforce. Hence the healthcare sector presents many varied opportunities for generative AI in research, clinical, operational, and behavioural applications. This chapter investigates the complex challenges in the social context of behavioural applications. In this domain, analytics examine large data sets (including large language models or LLMs) for client behaviour patterns that increase the probability of actions taken to improve engagement, well-being and health outcomes. Natural language processing is an emerging field that can offer psychological counselling and assistance for social and mental health both outside and within care facilities. Virtual and embodied bots have a growing role in the management of emotions, stress and outcomes at the interface between doctors, psychologists, nurses and patients. Yet at the current state of technology in both virtual and embodied socially assistive robots (SARs), the user interfaces with an artificial system that operates based on pre-trained databases which give only the temporary illusion of a relational agent. At present SARs are able to sustain only brief engagements, presenting major challenges for broader user-acceptance. Such databases may also contain situational, cultural, gender, racial and other biases. There are also significant transparency and privacy issues. The future is therefore not in general LLMs but in carefully trained and constantly-updating niche language models (NLMs). In designing and applying such vital systems all stakeholders, including regulatory bodies, medical professionals, nursing professionals and patient advocacy groups must be involved. Finally, assistive technologies must become part of the core curriculum of health professional education.  

When deep learning methods are applied to the detection of low contrast LCD surfaces, due to the imbalance between positive and negative samples and the difficulty in detecting micro defects with uneven brightness, we propose a method for automatic sample generation and detection based on deep generation network models. Firstly, generate several defect samples by generating adversarial networks to generate an expanded sample dataset. Secondly, the deep generation network is combined with the encoder to form an unsupervised model, and the defective parts of the image are obtained through image comparison. The experimental results confirm that the proposed method can automatically generate LCD image samples, and experiments on Mask R-CNN and unsupervised deep generation network models also confirm the effectiveness of our proposed method.

The rapid growth of digital pathology in recent years has provided an ideal opportunity for the development of artificial intelligence-based tools to improve the accuracy and efficiency of clinical diagnoses. One of the significant roadblocks to current research is the high level of visual variability across digital pathology images, causing models to generalise poorly to unseen data. Stain normalisation aims to standardise the visual profile of digital pathology images without changing the structural content of the images. In this chapter, we explore different techniques which have been used for stain normalisation in digital pathology, with a focus on approaches which utilise generative adversarial networks (GANs). Typically, GAN-based methods outperform non-generative approaches but at the cost of much greater computational requirements. However, it is not clear which method is best for stain normalisation in general, with different GAN and non-GAN approaches outperforming each other in different scenarios and according to different performance metrics. This is an ongoing field of study as researchers aim to identify a method which efficiently and effectively normalises pathology images to make AI models more robust and generalisable.

In many domains including medicine, biology, and neuroscience, rare events are the norm rather than the exception, limiting the ability to train intelligent systems to perform reliable pattern classification. This is the case when monitoring brain activity for epileptic seizures that constitute infrequent periods when abnormal electrical activity propagates across clusters of neurons. Here, as a solution, we describe how a generative adversarial network (GAN) can serve to produce synthetic examples that capture key features of epileptic activity observed in networks of in vitro cortical neurons. Further, GANs can generate novel patterns that deviate in systematic ways from the original data. A convolutional neural network whose goal was to classify healthy and seizure activity attained higher performance when trained on an augmented dataset composed of both original and synthetic data. Altogether, this work shows how GANs can provide data augmentation in a domain of epileptic seizures characterized by rare events.

Innovative strategies to promote social support and a sense of belonging are needed urgently as one in three adults worldwide experience loneliness. This Chapter explores the prospect of generative artificial intelligence (AI) chatbots in social support interventions to improve an individual's sense of belonging, social support, and reduce loneliness. This Chapter reviews the prominent areas that AI chatbots are currently being implemented and their outcomes. It compares AI chatbots designed for social companionship with those designed for assistance such as ChatGPT and Bard. It investigates individuals who are more vulnerable and susceptible to using AI chatbots and the possible positive outcomes and negative effects to autonomy. Ethical considerations and limitations of the integration of AI chatbots being employed into today’s society are debated especially in terms of loneliness. Together, the arguments in this Chapter propose the benefits of using AI chatbots as an assistive tool to improve overall well-being by managing time, advising, offering suggestions, and collaborating with the user to indirectly promote a sense of belonging and lessen feelings of loneliness.

The rapid advancement of generative artificial intelligence (AI) and large language model (LLM) technologies, such as ChatGPT-4 and Bard, has the potential to significantly change wellbeing education. This Chapter explores the applications of generative AI technologies in wellbeing education, with a focus on how chatbots and similar can be used to cultivate wellbeing through the SEARCH framework. For clarity, the SEARCH framework focuses on developing Strengths, Emotional management, Attention and awareness, Relationships, Coping, and Habits and goals. We begin by presenting the potential benefits of incorporating generative AI in wellbeing education. Next, by employing the SEARCH framework as a model of wellbeing education, the Chapter broadly conceptualises how AI technologies can be used to teach and explore the SEARCH components. The potential impact of AI-enhanced wellbeing education on teaching and learning practice—with implications for preparing teachers with ethical considerations and practical knowledge for using such technology—are also discussed.

The recent use of Generative AI (GenAI) techniques in Ecology has provided insights into predicting species co-occurrence patterns, specifically in water-limited ecosystems where multispecific plant clumps grow sparsely. In particular, these patterns have been employed to elucidate the mechanisms governing the assembly of plant communities in the context of Southeastern Spain. We discuss how the important concepts of transfer learning, and data augmentation take on slightly different meanings in this context, as compared to their usual application in Computer Vision. In particular, using transfer learning, the same models have been successfully applied to other plant communities in another semi-arid region of Spain and of tropical Mexico, opening the door to a specific kind of data augmentation by combining data sets from disparate communities. Beyond that, we also discuss the use of GenAI for synthetic data, and for predictions that can be of practical use when replanting vegetation in degraded environments, with an eye to biodiversity. 

Intelligent fault diagnosis often requires a balanced dataset which is hard to be obtained in industrial equipments, often resulting in an imbalance between data in normal and data in the presence of faults. Data augmentation techniques are among the most promising approaches to mitigate such issue. Generative adversarial networks (GAN) are a type of generative model consisting of a generator module and a discriminator. Through adversarial learning between these modules, the optimised generator can produce synthetic patterns that can be used for data augmentation. We investigate the role of loss function in improving the training efficiency of GAN. We proposed a generalization of both mean square error (MSE GAN) and Wasserstein GAN with gradient penalty (WGAN-GP), referred to as VGAN (the V-matrix based GAN) to mitigate training instability. Also, we investigate the sliced Wasserstein distance (SWD) as the loss function of a cycle consistency generative adversarial network (CycleGAN), referred to as SW-CycleGAN. Both two models are evaluated on an industrial robot data set. Experimental results show that the proposed loss functions outperform other competitive models especially in terms of computational costs.

This chapter introduces underwater acoustic noise modeling based on generative-adversarial-network (GAN). In underwater acoustic communications, accurately fitting the impulsive noise is crucial. Traditional models with fixed parameters can only approximate the global heavy-tail distribution of the impulsive noise, failing to capture local distributions of varying lengths. To address this limitation, a GAN-based underwater noise simulator (GANUNS) has been presented. The GANUNS consists of a deep-neural-network-based generator and a convolutional-neural-network-based discriminator that learn the heavy-tail distribution of the impulsive noise. By utilizing real noise data collected in Xiamen, the simulated underwater acoustic noise generated by the GANUNS exhibits significantly lower Kullback–Leibler divergence, Jensen-Shannon divergence, and mean square error compared to traditional approximate models. This demonstrates the effectiveness of the suggested GANUNS in accurately modeling the impulsive noise for underwater acoustic communications.

Medical imaging is a crucial aspect of modern healthcare, as it enables the diagnosis and treatment of various diseases and conditions. However, developing and deploying AI models for medical imaging is challenging, due to the limited availability and quality of data, as well as the high complexity and diversity of imaging modalities and tasks. Generative AI models, such as variational autoencoders (VAEs), generative adversarial networks (GANs), and text-to-image diffusion models, offer a promising solution to these challenges, as they can generate realistic and diverse images from existing data or latent representations. In this chapter, we provide a practical guide on how generative AI is transforming medical imaging, by reviewing the state-of-the-art methods and frameworks, presenting some successful case studies in different domains and modalities, and discussing the future directions and opportunities for research and development.

Deep learning techniques have made its way to the medical field. Medical images are essential tools for visualizing internal body structures up to cellular levels. X-ray computed tomography (CT) is a widely used non-invasive medical modality for patient diagnosis. Harmful effects of cumulative amounts of radiation exposure to patients undergoing CT scan have been recorded which includes hair loss, cancer and other illnesses. The “As Low as Reasonably Achievable” (ALARA) principle was developed with the purpose of minimizing the radiation dose to patients. This chapter discusses the implementation of artificial intelligence to devices for the reconstruction of CT images affected by the reduction of the radiation. The corrupted CT images have noticeable noise and artifacts that causes inaccuracies of medical diagnosis. One of the robust deep learning models for LDCT restoration is the Generative Adversarial Networks (GAN). This study shows a simple GAN architecture that aims to minimize edge over-smoothing, image texture enhancement and preservation of structural details of the medical images. Further, a benchmark testing was done to show the performance of the network compared with other state of the art models (SOTA). In addition, ablation experiments for the modules used in the network and loss functions used for the training procedure are also presented.

As the capacity for visual representation continues to evolve, there is a growing need for techniques for realistic and efficient creation of three-dimensional objects. Generative models, particularly Generative Adversarial Networks, Variational Autoencoders and novel methods of Text-to-3D, utilize textual descriptions to generate 3D reconstructions with high-quality geometry and disentangled materials. In this chapter, we present an in-depth exploration of the application of generative models in 3D reconstruction. We begin by discussing the theoretical underpinnings of these models and their applicability to 3D reconstruction. This chapter studies how these models learn to generate new instances from a given distribution. We end with discussions of potential future directions and the broader impacts of these technologies in various industries.

Novel artificial intelligence (AI) technologies have the capability to reinvent education and enhance the immersive experience of learners. The education industry is continually embracing advanced AI tools to harness their functionalities for optimal teaching and learning. When the metaverse was released, its interactive features were leveraged to promote an engaging learning experience for students. Several applications of the metaverse in schools indicate that metaverse-based education has the potential to provide situated and authentic learning which is highly prized in this twenty-first century. ChatGPT is one such AI tool which is touted to revolutionize education. ChatGPT 3.5 and 4 possess powerful functionalities that have been underscored to provide a fully immersive experience for learners. The quick response rate, level of accuracy, personality, and other abilities of ChatGPT makes it an enabler of metaverse-based education to create a higher level of immersiveness. In this review article, we outline the strengths and weaknesses of both the metaverse and ChatGPT and how ChatGPT can be used to overcome the shortcomings of metaverse-based education. Considerations for implementing ChatGPT in the metaverse are also discussed. It is proposed that educators should integrate ChatGPT into metaverse educational platforms to situate learners in realistic, problem-solving, and creative ways of thinking.

This chapter addresses generating artistic portrait drawings (APDrawings) from images, and we focus on two methods based on generative adversarial networks (GANs). We first introduce the genre of portrait line drawings, and review some existing methods for generating them from images. We also describe the Artistic Portrait Drawing (APDrawing) dataset, which contains 140 high-resolution face photos and corresponding portrait drawings executed by a professional artist. We then describe the APDrawingGAN method, which is a hierarchical GAN model that learns from paired data of face photos and portrait drawings, and the QMUPD method, which can learn from unpaired data of face photos and drawing. APDrawingGAN uses a novel distance transform loss to learn stroke lines in the drawings, and a local transfer loss to capture different drawing styles for different facial regions. QMUPD uses an asymmetric cycle mapping to preserve important facial features, and a quality metric to guide the generation towards high-quality drawings. We further introduce some recent developments which are based on multiple scale analysis, 3D information and multi-modal information. Finally, we describe the evaluation of artistic portrait drawings, which is a challenging task since there are many possible drawings that would be considered by experts to be acceptable.

Artificial intelligence (AI) deep learning generative models play an increasingly important role in drug design. Developments of different drug generative models can save capital and time to promote new drug discovery. AI deep learning generative models can be divided into different generative models based on the different levels of dimensional features of receptors and ligands such as SMILES generative models, molecular graph generative models, and 3D molecule generative models. Besides, based on the different algorithms, AI deep learning generative models for drug discovery can be roughly classified as variational autoencoder generative model, generative adversarial network generative model, and flow based generative model, and diffusion generative model. In this chapter, the classification, general mathematical methods, and research reports of AI deep learning generative models are summarized based on the different levels of dimensional features and algorithms. This chapter proposes an interesting topic and a deep understanding of AI deep learning generative models for the scientific community.

The area of 3D generative network has been developing rapidly due in part to the progresses in generative models and 3D sensing technology. It has a wide range of applications in film and animation, video games, virtual reality, etc. Although many popular 3D generative networks are inspired from the 2D ones, they are significantly different. This is essentially because the representations of 3D data, particularly the non-Euclidean ones, cannot be directly processed by 2D generative networks. In this chapter, we first present an overview of 3D generative networks. Then, we introduce the common representations of 3D data, including Euclidean and non-Euclidean ones. Next, we present the mainstream methods for 3D generative networks categorised subject to the same taxonomy as the 2D generative models. Finally, we discuss the limitations of 3D generative networks and potential future work in this field.

The chapter focuses on the economic aspects of generative AI. It looks at the cost–benefit analysis of generative AI implementation in a company, delves into the automatability of tasks by generative AI, evaluates its substitution, enhancement (augmentation) and transformational effects, discusses it economic limitations and its micro-and macroeconomic implications. The analysis will be helpful to managers and owners who consider the adoption of generative AI by their organisations.

Plant phenotyping is the study of plants’ physiological, morphological and biochemical traits resulting from their interaction with the environment. These traits (e.g., leaf area, leaf count, tillering, wilting etc.) are crucial in current plant research, focused on improving plant quality i.e., disease resistance, drought resistance and productivity. With the advancement in sensor technologies, image based analysis via various computer vision methods (e.g., image classification, segmentation, object detection etc.) have emerged in plant phenotyping. Specifically, state-of-the-art deep learning models have been employed for high-throughput study of plant traits. However, the application of deep learning models is currently limited due to the high variability in plant traits among various plant species and unstructured plant imaging. Additionally, complex plant traits pose high data collection and annotation costs. In this context, generative artificial intelligence (AI) based on the evolution of generative adversarial networks (GANs) for data synthesis can relieve the current bottleneck of data scarcity and plant species gap. This chapter reviews the application of state-of-the-art GANs for plant image datasets such as leaf, weed, disease etc. It also discusses the current Generative AI challenges and future directions for agriculture data synthesis.

Missing data poses an ubiquitous challenge across a wide range of applications, stemming from a multitude of causes that are both diverse and context-dependent. The prevailing issue is that most advanced data analysis techniques are primarily tailored for complete datasets, thereby underscoring the indispensable need for effective imputation methods. In this chapter, we embark on an extensive exploration of missing data from a statistical perspective, offering a holistic review of its intricate nature. Our investigation encompasses a deep dive into the various mechanisms underlying missing data, shedding light on their ignorability and identifiability-fundamental concepts essential for understanding and addressing this pervasive issue. Moreover, we present a succinct yet comprehensive overview of influential classical imputation methods, showcasing their contributions to the field. Building upon this foundation, we delve into the latest advancements in generative models, a burgeoning area that holds great promise for learning from and imputing missing data. By harnessing the power of generative models, we aim to unlock novel insights and methodologies that can tackle the challenges posed by missing data. Furthermore, we introduce an approach that specifically addresses the critical problem of nonparametric identifiability in nonignorable missing data through the innovative use of generative models. This novel approach aims to overcome the limitations associated with alternative generative models and provides a potential solution to this challenging issue. To enhance the clarity of our proposed method, we supplement our discourse with curated numerical examples that distinguish its effectiveness from other baselines in specific scenarios. Through the exploration, we hope to pave the way for further research and advancements in this critical domain, ultimately leading to more accurate and reliable analyses and interpretations of incomplete datasets.

The ability of generative models to accurately fit data distributions has resulted in their widespread adoption and success in fields such as computer vision and natural language processing. In this chapter, we provide a brief overview of the application of generative models in the domain of infrared (IR) image super-resolution, including a discussion of the various challenges and adversarial training methods employed. We propose potential areas for further investigation and advancement in the application of generative models for IR image super-resolution.

In the field of fire safety, Generative AI presents promising opportunities to enhance prevention, response, and recovery efforts. In this chapter, we explore the potential applications of Generative AI in fire safety. Generative AI benefits the fire safety field in applications including fires simulation and emergency response training, predictive analytics for fire detection and prediction, evacuation planning and optimization, firefighting robotics, and post-fire reconstruction as well as fire investigation. In this chapter, we provided details of two empirical vision-based examples of employing Generative AI for fire safety applications. These examples show how powerful solutions Generative AI models could be for cases where data shortages have been a hurdle for the advancement of AI for fire safety. With all benefits that Generative AI provides us, careful testing, adherence to safety standards, and collaboration between AI experts and fire safety professionals are crucial to ensure the responsible and effective implementation of Generative AI in the context of fire safety.

In recent years, intelligent fault diagnosis based on deep learning has achieved vigorous development due to its powerful feature representation ability. However, the data collected from industrial sites often contain different levels of noise, which makes it difficult to extract effective fault features, which seriously affects the performance of the model. In addition, the limited labeled data makes the training of deep network models even more challenging. To address the above problems, a multi-scale convolutional autoencoder with attention mechanism (MSCAE-AM) is developed. Specifically, as a typical unsupervised learning model, the encoder can effectively reduce the dependence on labeled data. Furthermore, the feature extraction ability of the model in noisy environments can be improved by combining noise reduction operations and embedding multi-scale convolutional layers and attention mechanisms. Experimental results on the wind turbine fault simulation datasets verify the effectiveness and superiority of the proposed method. The results show that the proposed method can not only effectively reduce the dependence on label data, but also has stronger robustness to noise than other methods.