In this work, we propose two convolutional neural network (CNN) based metrics, classification score (CS) and distribution score (DS), for performance evaluation of generative adversarial networks (GANs). Though GAN-generated images can be evaluated through manual assessment of visual fidelity, it is prolonged, subjective, challenging, tiresome, and can be misleading. Existing quantitative methods are biased toward memory GAN and fail to detect overfitting. CS and DS allow us to experimentally prove that training of GANs is actually guided by the dataset that it improves with every epoch and gets closer to following the distribution of the dataset. Both methods are based on GAN-generated image classification by CNN. CS is the root mean square (rms) value of three different classification techniques, direct classification (DC), indirect classification (IC), and blind classification (BC). It exhibits the degree to which GAN can learn the features and generate fake images similar to real datasets. DS shows the contrast between the mean distribution of GAN-generated data and the real data. It indicates the extent to which GANs can create synthetic images with similar distribution to real datasets. We evaluated CS and DS metrics for different variants of GANs and compared their performances with existing metrics. Results show that CS and DS can evaluate the different variants of GANs quantitatively and qualitatively while detecting overfitting and mode collapse.
在这项工作中，我们提出了两个基于卷积神经网络（CNN）的指标，即分类得分（CS）和分布得分（DS），用于生成式对抗网络（GAN）的性能评估。虽然可以通过人工评估视觉保真度来评价 GAN 生成的图像，但这种方法耗时长、主观性强、具有挑战性、令人厌烦，而且可能会产生误导。现有的量化方法偏重于记忆 GAN，无法检测到过拟合。CS 和 DS 可以让我们通过实验证明，GAN 的训练实际上是在数据集的指导下进行的，它在每个历时中都会有所改进，并更接近于遵循数据集的分布。这两种方法都是基于由 CNN 生成的 GAN 图像分类。CS 是三种不同分类技术（直接分类 (DC)、间接分类 (IC) 和盲分类 (BC)）的均方根值。它显示了 GAN 学习特征并生成与真实数据集相似的假图像的程度。DS 显示了 GAN 生成数据的平均分布与真实数据之间的对比。它表明 GAN 能够生成与真实数据集分布相似的合成图像的程度。我们针对 GAN 的不同变体评估了 CS 和 DS 指标，并将其性能与现有指标进行了比较。结果表明，CS 和 DS 可以定量和定性地评估 GAN 的不同变体，同时检测过度拟合和模式崩溃。
{"title":"CNN-Based Metrics for Performance Evaluation of Generative Adversarial Networks","authors":"Adarsh Prasad Behera;Satya Prakash;Siddhant Khanna;Shivangi Nigam;Shekhar Verma","doi":"10.1109/TAI.2024.3401650","DOIUrl":"https://doi.org/10.1109/TAI.2024.3401650","url":null,"abstract":"In this work, we propose two convolutional neural network (CNN) based metrics, classification score (CS) and distribution score (DS), for performance evaluation of generative adversarial networks (GANs). Though GAN-generated images can be evaluated through manual assessment of visual fidelity, it is prolonged, subjective, challenging, tiresome, and can be misleading. Existing quantitative methods are biased toward memory GAN and fail to detect overfitting. CS and DS allow us to experimentally prove that training of GANs is actually guided by the dataset that it improves with every epoch and gets closer to following the distribution of the dataset. Both methods are based on GAN-generated image classification by CNN. CS is the root mean square (rms) value of three different classification techniques, direct classification (DC), indirect classification (IC), and blind classification (BC). It exhibits the degree to which GAN can learn the features and generate fake images similar to real datasets. DS shows the contrast between the mean distribution of GAN-generated data and the real data. It indicates the extent to which GANs can create synthetic images with similar distribution to real datasets. We evaluated CS and DS metrics for different variants of GANs and compared their performances with existing metrics. Results show that CS and DS can evaluate the different variants of GANs quantitatively and qualitatively while detecting overfitting and mode collapse.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443099","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Existing multiscale methods lead to a risk of just increasing the receptive field sizes while neglecting small receptive fields. Thus, it is a challenging problem to effectively construct adaptive neural networks for recognizing various spatial-scale objects. To tackle this issue, we first introduce a new attention dimension, i.e., depth, in addition to existing attentions such as channel-attention, spatial-attention, branch-attention, and self-attention. We present a novel selective depth attention network to treat multiscale objects symmetrically in various vision tasks. Specifically, the blocks within each stage of neural networks, including convolutional neural networks (CNNs), e.g., ResNet, SENet, and Res2Net, and vision transformers (ViTs), e.g., PVTv2, output the hierarchical feature maps with the same resolution but different receptive field sizes. Based on this structural property, we design a depthwise building module, namely an selective depth attention (SDA) module, including a trunk branch and a SE-like attention branch. The block outputs of the trunk branch are fused to guide their depth attention allocation through the attention branch globally. According to the proposed attention mechanism, we dynamically select different depth features, which contributes to adaptively adjusting the receptive field sizes for the variable-sized input objects. Moreover, our method is orthogonal to multiscale networks and attention networks, so-called SDA- $x$ Net. Extensive experiments demonstrate that the proposed SDA method significantly improves the original performance as a lightweight and efficient plug-in on numerous computer vision tasks, e.g., image classification, object detection, and instance segmentation.
{"title":"Selective Depth Attention Networks for Adaptive Multiscale Feature Representation","authors":"Qingbei Guo;Xiao-Jun Wu;Tianyang Xu;Tongzhen Si;Cong Hu;Jinglan Tian","doi":"10.1109/TAI.2024.3401652","DOIUrl":"https://doi.org/10.1109/TAI.2024.3401652","url":null,"abstract":"Existing multiscale methods lead to a risk of just increasing the receptive field sizes while neglecting small receptive fields. Thus, it is a challenging problem to effectively construct adaptive neural networks for recognizing various spatial-scale objects. To tackle this issue, we first introduce a new attention dimension, i.e., depth, in addition to existing attentions such as channel-attention, spatial-attention, branch-attention, and self-attention. We present a novel selective depth attention network to treat multiscale objects symmetrically in various vision tasks. Specifically, the blocks within each stage of neural networks, including convolutional neural networks (CNNs), e.g., ResNet, SENet, and Res2Net, and vision transformers (ViTs), e.g., PVTv2, output the hierarchical feature maps with the same resolution but different receptive field sizes. Based on this structural property, we design a depthwise building module, namely an selective depth attention (SDA) module, including a trunk branch and a SE-like attention branch. The block outputs of the trunk branch are fused to guide their depth attention allocation through the attention branch globally. According to the proposed attention mechanism, we dynamically select different depth features, which contributes to adaptively adjusting the receptive field sizes for the variable-sized input objects. Moreover, our method is orthogonal to multiscale networks and attention networks, so-called SDA-\u0000<inline-formula><tex-math>$x$</tex-math></inline-formula>\u0000Net. Extensive experiments demonstrate that the proposed SDA method significantly improves the original performance as a lightweight and efficient plug-in on numerous computer vision tasks, e.g., image classification, object detection, and instance segmentation.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142442955","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-15DOI: 10.1109/TAI.2024.3401649
Shaoqi Sun;Kele Xu;Dawei Feng;Bo Ding
In recent years, multiagent reinforcement learning (MARL) has demonstrated considerable potential across diverse applications. However, in reinforcement learning environments characterized by sparse rewards, the scarcity of reward signals may give rise to reward conflicts among agents. In these scenarios, each agent tends to compete to obtain limited rewards, deviating from collaborative efforts aimed at achieving collective team objectives. This not only amplifies the learning challenge but also imposes constraints on the overall learning performance of agents, ultimately compromising the attainment of team goals. To mitigate the conflicting competition for rewards among agents in MARL, we introduce the bidirectional influence and interaction (BDII) MARL framework. This innovative approach draws inspiration from the collaborative ethos observed in human social cooperation, specifically the concept of “sharing joys and sorrows.” The fundamental concept behind BDII is to empower agents to share their individual rewards with collaborators, fostering a cooperative rather than competitive behavioral paradigm. This strategic shift aims to resolve the pervasive issue of reward conflicts among agents operating in sparse-reward environments. BDII incorporates two key factors—namely, the Gaussian kernel distance between agents (physical distance) and policy diversity among agents (logical distance). The two factor collectively contribute to the dynamic adjustment of reward allocation coefficients, culminating in the formation of reward distribution weights. The incorporation of these weights facilitates the equitable sharing of agents’ contributions to rewards, promoting a cooperative learning environment. Through extensive experimental evaluations, we substantiate the efficacy of BDII in addressing the challenge of reward conflicts in MARL. Our research findings affirm that BDII significantly mitigates reward conflicts, ensuring that agents consistently align with the original team objectives, thereby achieving state-of-the-art performance. This validation underscores the potential of the proposed framework in enhancing the collaborative nature of multiagent systems, offering a promising avenue for advancing the field of reinforcement learning.
{"title":"Bidirectional Influence and Interaction for Multiagent Reinforcement Learning","authors":"Shaoqi Sun;Kele Xu;Dawei Feng;Bo Ding","doi":"10.1109/TAI.2024.3401649","DOIUrl":"https://doi.org/10.1109/TAI.2024.3401649","url":null,"abstract":"In recent years, multiagent reinforcement learning (MARL) has demonstrated considerable potential across diverse applications. However, in reinforcement learning environments characterized by sparse rewards, the scarcity of reward signals may give rise to reward conflicts among agents. In these scenarios, each agent tends to compete to obtain limited rewards, deviating from collaborative efforts aimed at achieving collective team objectives. This not only amplifies the learning challenge but also imposes constraints on the overall learning performance of agents, ultimately compromising the attainment of team goals. To mitigate the conflicting competition for rewards among agents in MARL, we introduce the bidirectional influence and interaction (BDII) MARL framework. This innovative approach draws inspiration from the collaborative ethos observed in human social cooperation, specifically the concept of “sharing joys and sorrows.” The fundamental concept behind BDII is to empower agents to share their individual rewards with collaborators, fostering a cooperative rather than competitive behavioral paradigm. This strategic shift aims to resolve the pervasive issue of reward conflicts among agents operating in sparse-reward environments. BDII incorporates two key factors—namely, the Gaussian kernel distance between agents (physical distance) and policy diversity among agents (logical distance). The two factor collectively contribute to the dynamic adjustment of reward allocation coefficients, culminating in the formation of reward distribution weights. The incorporation of these weights facilitates the equitable sharing of agents’ contributions to rewards, promoting a cooperative learning environment. Through extensive experimental evaluations, we substantiate the efficacy of BDII in addressing the challenge of reward conflicts in MARL. Our research findings affirm that BDII significantly mitigates reward conflicts, ensuring that agents consistently align with the original team objectives, thereby achieving state-of-the-art performance. This validation underscores the potential of the proposed framework in enhancing the collaborative nature of multiagent systems, offering a promising avenue for advancing the field of reinforcement learning.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443126","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
It is not uncommon that real-world data are distributed with a long tail. For such data, the learning of deep neural networks becomes challenging because it is hard to classify tail classes correctly. In the literature, several existing methods have addressed this problem by reducing classifier bias, provided that the features obtained with long-tailed data are representative enough. However, we find that training directly on long-tailed data leads to uneven embedding space. That is, the embedding space of head classes severely compresses that of tail classes, which is not conducive to subsequent classifier learning. This article therefore studies the problem of long-tailed visual recognition from the perspective of feature level. We introduce feature augmentation to balance the embedding distribution. The features of different classes are perturbed with varying amplitudes in Gaussian form. Based on these perturbed features, two novel logit adjustment methods are proposed to improve model performance at a modest computational overhead. Subsequently, the distorted embedding spaces of all classes can be calibrated. In such balanced-distributed embedding spaces, the biased classifier can be eliminated by simply retraining the classifier with class-balanced sampling data. Extensive experiments conducted on benchmark datasets demonstrate the superior performance of the proposed method over the state-of-the-art ones.
{"title":"Adjusting Logit in Gaussian Form for Long-Tailed Visual Recognition","authors":"Mengke Li;Yiu-ming Cheung;Yang Lu;Zhikai Hu;Weichao Lan;Hui Huang","doi":"10.1109/TAI.2024.3401102","DOIUrl":"https://doi.org/10.1109/TAI.2024.3401102","url":null,"abstract":"It is not uncommon that real-world data are distributed with a long tail. For such data, the learning of deep neural networks becomes challenging because it is hard to classify tail classes correctly. In the literature, several existing methods have addressed this problem by reducing classifier bias, provided that the features obtained with long-tailed data are representative enough. However, we find that training directly on long-tailed data leads to uneven embedding space. That is, the embedding space of head classes severely compresses that of tail classes, which is not conducive to subsequent classifier learning. This article therefore studies the problem of long-tailed visual recognition from the perspective of feature level. We introduce feature augmentation to balance the embedding distribution. The features of different classes are perturbed with varying amplitudes in Gaussian form. Based on these perturbed features, two novel logit adjustment methods are proposed to improve model performance at a modest computational overhead. Subsequently, the distorted embedding spaces of all classes can be calibrated. In such balanced-distributed embedding spaces, the biased classifier can be eliminated by simply retraining the classifier with class-balanced sampling data. Extensive experiments conducted on benchmark datasets demonstrate the superior performance of the proposed method over the state-of-the-art ones.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10531112","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443091","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-14DOI: 10.1109/TAI.2024.3375833
Jiang-Wen Xiao;Hongliang Fang;Yan-Wu Wang
Short-term residential load forecasting is essential to demand side response. However, the frequent spikes in the load and the volatile daily load patterns make it difficult to accurately forecast the load. To deal with these problems, this article proposes a smoothing clustering method for daily load clustering and a pooling-ensemble model for one day ahead load forecasting. The whole short-term load forecasting framework in this article contains three steps. Specifically and first, the states of the residents are obtained by clustering the daily load curves with the proposed smoothing clustering method. Second, a weighted mixed Markov model is built to predict the probability distribution of the load state in the next day. Third, multiple predictors in the pooling-ensemble model are selected for different states and the load is forecasted by weighing the results of the multiple predictors based on the predicted states. Results of the case studies and comparison studies on two public datasets verify the advantages of the smoothing clustering method and the pooling-ensemble model.
{"title":"Short-Term Residential Load Forecasting via Pooling-Ensemble Model With Smoothing Clustering","authors":"Jiang-Wen Xiao;Hongliang Fang;Yan-Wu Wang","doi":"10.1109/TAI.2024.3375833","DOIUrl":"https://doi.org/10.1109/TAI.2024.3375833","url":null,"abstract":"Short-term residential load forecasting is essential to demand side response. However, the frequent spikes in the load and the volatile daily load patterns make it difficult to accurately forecast the load. To deal with these problems, this article proposes a smoothing clustering method for daily load clustering and a pooling-ensemble model for one day ahead load forecasting. The whole short-term load forecasting framework in this article contains three steps. Specifically and first, the states of the residents are obtained by clustering the daily load curves with the proposed smoothing clustering method. Second, a weighted mixed Markov model is built to predict the probability distribution of the load state in the next day. Third, multiple predictors in the pooling-ensemble model are selected for different states and the load is forecasted by weighing the results of the multiple predictors based on the predicted states. Results of the case studies and comparison studies on two public datasets verify the advantages of the smoothing clustering method and the pooling-ensemble model.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141630935","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-14DOI: 10.1109/TAI.2024.3377147
Tamador Mohaidat;Kasem Khalil
Artificial intelligence (AI) hardware accelerator is an emerging research for several applications and domains. The hardware accelerator's direction is to provide high computational speed with retaining low-cost and high learning performance. The main challenge is to design complex machine learning models on hardware with high performance. This article presents a thorough investigation into machine learning accelerators and associated challenges. It describes a hardware implementation of different structures such as convolutional neural network (CNN), recurrent neural network (RNN), and artificial neural network (ANN). The challenges such as speed, area, resource consumption, and throughput are discussed. It also presents a comparison between the existing hardware design. Last, the article describes the evaluation parameters for a machine learning accelerator in terms of learning and testing performance and hardware design.
{"title":"A Survey on Neural Network Hardware Accelerators","authors":"Tamador Mohaidat;Kasem Khalil","doi":"10.1109/TAI.2024.3377147","DOIUrl":"https://doi.org/10.1109/TAI.2024.3377147","url":null,"abstract":"Artificial intelligence (AI) hardware accelerator is an emerging research for several applications and domains. The hardware accelerator's direction is to provide high computational speed with retaining low-cost and high learning performance. The main challenge is to design complex machine learning models on hardware with high performance. This article presents a thorough investigation into machine learning accelerators and associated challenges. It describes a hardware implementation of different structures such as convolutional neural network (CNN), recurrent neural network (RNN), and artificial neural network (ANN). The challenges such as speed, area, resource consumption, and throughput are discussed. It also presents a comparison between the existing hardware design. Last, the article describes the evaluation parameters for a machine learning accelerator in terms of learning and testing performance and hardware design.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141980079","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-12DOI: 10.1109/TAI.2024.3376640
Sudipta Bhuyan;Aupendu Kar;Debashis Sen;Sankha Deb
Significant improvement has been achieved lately in color and depth data-based salient object detection (SOD) on images from varied datasets, which is mainly due to RGB-D fusion using modern machine learning techniques. However, little emphasis has been given recently on performing RGB-D fusion for SOD in the absence of ground truth data for training. This article proposes a zero-shot deep RGB-D fusion approach based on the novel concept of fuzzy membership learning, which does not require any data for training. The constituent salient object maps to be fused are represented using parametric fuzzy membership functions and the optimal parameter values are estimated through our zero-shot fuzzy membership learning (Z-FML) network. The optimal parameter values are used in a fuzzy inference system along with the constituent salient object maps to perform the fusion. A measure called the membership similarity measure (MSM) is proposed, and the Z-FML network is trained using it to devise a loss function that maximizes the similarity between the constituent salient object maps and the fused salient object map. The deduction of MSM and its properties are shown theoretically, and the gradients involved in the training of the Z-FML network are derived. Qualitative and quantitative evaluations using several datasets signify the effectiveness of our RGB-D fusion and our fusion-based RGB-D SOD in comparison with the state-of-the-art. We also empirically demonstrate the advantage of employing the novel MSM for training our Z-FML network.
{"title":"RGB-D Fusion Through Zero-Shot Fuzzy Membership Learning for Salient Object Detection","authors":"Sudipta Bhuyan;Aupendu Kar;Debashis Sen;Sankha Deb","doi":"10.1109/TAI.2024.3376640","DOIUrl":"https://doi.org/10.1109/TAI.2024.3376640","url":null,"abstract":"Significant improvement has been achieved lately in color and depth data-based salient object detection (SOD) on images from varied datasets, which is mainly due to RGB-D fusion using modern machine learning techniques. However, little emphasis has been given recently on performing RGB-D fusion for SOD in the absence of ground truth data for training. This article proposes a zero-shot deep RGB-D fusion approach based on the novel concept of fuzzy membership learning, which does not require any data for training. The constituent salient object maps to be fused are represented using parametric fuzzy membership functions and the optimal parameter values are estimated through our zero-shot fuzzy membership learning (Z-FML) network. The optimal parameter values are used in a fuzzy inference system along with the constituent salient object maps to perform the fusion. A measure called the membership similarity measure (MSM) is proposed, and the Z-FML network is trained using it to devise a loss function that maximizes the similarity between the constituent salient object maps and the fused salient object map. The deduction of MSM and its properties are shown theoretically, and the gradients involved in the training of the Z-FML network are derived. Qualitative and quantitative evaluations using several datasets signify the effectiveness of our RGB-D fusion and our fusion-based RGB-D SOD in comparison with the state-of-the-art. We also empirically demonstrate the advantage of employing the novel MSM for training our Z-FML network.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141630936","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-11DOI: 10.1109/TAI.2024.3375832
Praveer Saxena;Ashish Kumar Bhandari
Colorectal cancer stands out as a major factor in cancer-related fatalities. The prevention of colorectal cancer may be aided by early polyp diagnosis. Colonoscopy is a widely used procedure for the diagnosis of polyps, but it is highly dependent on the skills of the medical practitioner. Automatic polyp segmentation using computer-aided diagnosis can help medical practitioners detect even those polyps missed by humans, and this early detection of polyps can save precious human lives. Due to the lack of distinct edges, poor contrast between the foreground and background, and great variety of polyps, automatic segmentation of polyps is quite difficult. Although there are several deep learning-based strategies for segmenting polyps, typical convolutional neural network (CNN)-based algorithms lack long-range dependencies and lose spatial information because of consecutive convolution and pooling. In this research, a novel encoder–decoder-based segmentation architecture has been proposed in an effort to identify distinguishing features that can be used to precisely separate the polyps. The proposed architecture combines the strengths of a pretrained ResNet50 encoder, residual block, our proposed multiscale dilation block, and the mask attention block. Multiscale dilation block enables us to extract features at different scales for better feature representation. The mask attention block utilizes a generated auxiliary mask in order to concentrate on important image features. To evaluate the proposed architecture, several polyp segmentation datasets have been used. The obtained findings show that the suggested architecture performs better than several state-of-the-art (SOTA) approaches for segmenting the polyps.
{"title":"Context Aware Automatic Polyp Segmentation Network With Mask Attention","authors":"Praveer Saxena;Ashish Kumar Bhandari","doi":"10.1109/TAI.2024.3375832","DOIUrl":"10.1109/TAI.2024.3375832","url":null,"abstract":"Colorectal cancer stands out as a major factor in cancer-related fatalities. The prevention of colorectal cancer may be aided by early polyp diagnosis. Colonoscopy is a widely used procedure for the diagnosis of polyps, but it is highly dependent on the skills of the medical practitioner. Automatic polyp segmentation using computer-aided diagnosis can help medical practitioners detect even those polyps missed by humans, and this early detection of polyps can save precious human lives. Due to the lack of distinct edges, poor contrast between the foreground and background, and great variety of polyps, automatic segmentation of polyps is quite difficult. Although there are several deep learning-based strategies for segmenting polyps, typical convolutional neural network (CNN)-based algorithms lack long-range dependencies and lose spatial information because of consecutive convolution and pooling. In this research, a novel encoder–decoder-based segmentation architecture has been proposed in an effort to identify distinguishing features that can be used to precisely separate the polyps. The proposed architecture combines the strengths of a pretrained ResNet50 encoder, residual block, our proposed multiscale dilation block, and the mask attention block. Multiscale dilation block enables us to extract features at different scales for better feature representation. The mask attention block utilizes a generated auxiliary mask in order to concentrate on important image features. To evaluate the proposed architecture, several polyp segmentation datasets have been used. The obtained findings show that the suggested architecture performs better than several state-of-the-art (SOTA) approaches for segmenting the polyps.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141706608","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-11DOI: 10.1109/TAI.2024.3375834
Prabuddha Chakraborty;Swarup Bhunia
Edge computing devices in Internet-of-Things (IoT) systems are being widely used in diverse application domains including industrial automation, surveillance, and smart housing. These applications typically employ a large array of sensors, store a high volume of data, and search within the stored data for specific patterns using machine intelligence. Due to this heavy reliance on data in these applications, optimizing the memory performance in edge devices has become an important research focus. In this work, we note (based on some preliminary quantitative studies) that the memory requirements of such application-specific systems tend to differ drastically from traditional general-purpose computing systems. Inspired by these findings and also through drawing inspiration from the human brain (which excels at being highly adaptive), we design a digital memory framework that can continually adapt to the specific needs of different edge devices. This adaption is made possible through a continual reinforcement-based learning methodology, and it aims at creating a digital memory framework that is always self-aware of the data it hold and queries being made. Through a methodical implementation of the framework, we demonstrate its effectiveness for different use-cases, settings, and hyperparameters in comparison with traditional content-addressable memory.
{"title":"A Self-Aware Digital Memory Framework Powered by Artificial Intelligence","authors":"Prabuddha Chakraborty;Swarup Bhunia","doi":"10.1109/TAI.2024.3375834","DOIUrl":"https://doi.org/10.1109/TAI.2024.3375834","url":null,"abstract":"Edge computing devices in Internet-of-Things (IoT) systems are being widely used in diverse application domains including industrial automation, surveillance, and smart housing. These applications typically employ a large array of sensors, store a high volume of data, and search within the stored data for specific patterns using machine intelligence. Due to this heavy reliance on data in these applications, optimizing the memory performance in edge devices has become an important research focus. In this work, we note (based on some preliminary quantitative studies) that the memory requirements of such application-specific systems tend to differ drastically from traditional general-purpose computing systems. Inspired by these findings and also through drawing inspiration from the human brain (which excels at being highly adaptive), we design a digital memory framework that can continually adapt to the specific needs of different edge devices. This adaption is made possible through a continual reinforcement-based learning methodology, and it aims at creating a digital memory framework that is always self-aware of the data it hold and queries being made. Through a methodical implementation of the framework, we demonstrate its effectiveness for different use-cases, settings, and hyperparameters in comparison with traditional content-addressable memory.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141630971","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-11DOI: 10.1109/TAI.2024.3375828
Chengyuan Yan;Jianwei Xia;Ju H. Park;Xiangpeng Xie
In this note, an adaptive event-triggered optimized tracking control problem is investigated for nonlinear switched systems with flexible output constraints under extended mode-dependent average dwell time (MDADT). Initially, a new shifting function and an improved barrier function are constructed to solve flexible output constraints under the practical background. Subsequently, a global performance function with the exponential discount factor based on the error variable is designed under optimized backstepping (OB), which not only ensures that the performance function converges, but also evaluates the tracking performance of the system, reflecting the energy consumption. The corresponding Hamilton–Jacobi–Bellman (HJB) equation is constructed to solve the optimal control strategy. To remove the restriction on the maximum asynchronous time, an event-triggered optimization strategy for subsystems is utilized to exclude Zeno behavior. Furthermore, we demonstrate that the signals of the closed-loop system are bounded and the flexible output constraints are strictly obeyed. Finally, the application of the above control technique to the manipulator system is validated.
{"title":"Flexible Constraints-Based Adaptive Intelligent Event-Triggered Control for Slowly Switched Nonlinear Systems Using Reinforcement Learning","authors":"Chengyuan Yan;Jianwei Xia;Ju H. Park;Xiangpeng Xie","doi":"10.1109/TAI.2024.3375828","DOIUrl":"https://doi.org/10.1109/TAI.2024.3375828","url":null,"abstract":"In this note, an adaptive event-triggered optimized tracking control problem is investigated for nonlinear switched systems with flexible output constraints under extended mode-dependent average dwell time (MDADT). Initially, a new shifting function and an improved barrier function are constructed to solve flexible output constraints under the practical background. Subsequently, a global performance function with the exponential discount factor based on the error variable is designed under optimized backstepping (OB), which not only ensures that the performance function converges, but also evaluates the tracking performance of the system, reflecting the energy consumption. The corresponding Hamilton–Jacobi–Bellman (HJB) equation is constructed to solve the optimal control strategy. To remove the restriction on the maximum asynchronous time, an event-triggered optimization strategy for subsystems is utilized to exclude Zeno behavior. Furthermore, we demonstrate that the signals of the closed-loop system are bounded and the flexible output constraints are strictly obeyed. Finally, the application of the above control technique to the manipulator system is validated.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141630904","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：