Pub Date : 2024-03-12DOI: 10.1109/TETCI.2024.3369849
Qian Tang;Yiji Zhao;Hao Wu;Lei Zhang
Contrastive representation learning has been widely employed in attributed graph clustering and has demonstrated significant success. However, these methods have two problems: 1)According to an assumption that clusters are formed around a minority of central anchor nodes, the contrastive relationships between these anchors are not explored in previous works. 2)They fail to deal with biased sample pairs, which may degrade the representation quality and cause poor clustering performance. To solve the problems, we propose a framework termed GE-S-D for both node representation learning and clustering, which consists of an anchor sampling strategy, a low-pass graph encoder, and a debiasing strategy. Specifically, to reveal the contrastive relationships between anchors, we design a sampling strategy to select a small number of anchors and then construct a training set of positive and negative sample pairs for contrastive learning. Then, we introduce a low-pass graph encoder to propagate contrastive messages to all nodes and learn cluster-friendly node representations. Furthermore, to alleviate the interference of biased sample pairs, we design a debiasing strategy using K-Means on the node representations to obtain the clustering information and remove the false positive and false negative sample pairs in the training set for improving contrastive learning. The clustering performance is verified on five benchmark datasets, and our method is superior to many state-of-the-art methods according to quantitive and qualitative analysis.
{"title":"Node Clustering on Attributed Graph Using Anchor Sampling Strategy and Debiasing Strategy","authors":"Qian Tang;Yiji Zhao;Hao Wu;Lei Zhang","doi":"10.1109/TETCI.2024.3369849","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3369849","url":null,"abstract":"Contrastive representation learning has been widely employed in attributed graph clustering and has demonstrated significant success. However, these methods have two problems: 1)According to an assumption that clusters are formed around a minority of central anchor nodes, the contrastive relationships between these anchors are not explored in previous works. 2)They fail to deal with biased sample pairs, which may degrade the representation quality and cause poor clustering performance. To solve the problems, we propose a framework termed GE-S-D for both node representation learning and clustering, which consists of an anchor sampling strategy, a low-pass graph encoder, and a debiasing strategy. Specifically, to reveal the contrastive relationships between anchors, we design a sampling strategy to select a small number of anchors and then construct a training set of positive and negative sample pairs for contrastive learning. Then, we introduce a low-pass graph encoder to propagate contrastive messages to all nodes and learn cluster-friendly node representations. Furthermore, to alleviate the interference of biased sample pairs, we design a debiasing strategy using K-Means on the node representations to obtain the clustering information and remove the false positive and false negative sample pairs in the training set for improving contrastive learning. The clustering performance is verified on five benchmark datasets, and our method is superior to many state-of-the-art methods according to quantitive and qualitative analysis.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141964860","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","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/TETCI.2024.3369321
Kui Jiang;Qiong Wang;Zhaoyi An;Zheng Wang;Cong Zhang;Chia-Wen Lin
Images captured in low-light or underwater environments are often accompanied by significant degradation, which can negatively impact the quality and performance of downstream tasks. While convolutional neural networks (CNNs) and Transformer architectures have made significant progress in computer vision tasks, there are few efforts to harmonize them into a more concise framework for enhancing such images. To this end, this study proposes to aggregate the individual capability of self-attention (SA) and CNNs for accurate perturbation removal while preserving background contents. Based on this, we carry forward a Retinex-based framework, dubbed as Mutual Retinex, where a two-branch structure is designed to characterize the specific knowledge of reflectance and illumination components while removing the perturbation. To maximize its potential, Mutual Retinex is equipped with a new mutual learning mechanism, involving an elaborately designed mutual representation module (MRM). In MRM, the complementary information between reflectance and illumination components are encoded and used to refine each other. Through the complementary learning via the mutual representation, the enhanced results generated by our model exhibit superior color consistency and naturalness. Extensive experiments have shown the significant superiority of our mutual learning based method over thirteen competitors on the low-light task and ten methods on the underwater image enhancement task. In particular, our proposed Mutual Retinex respectively surpasses the state-of-the-art method MIRNet-v2 by 0.90 dB and 2.46 dB in PSNR on the LOL 1000 and FIVEK datasets, while with only 19.8% model parameters.
{"title":"Mutual Retinex: Combining Transformer and CNN for Image Enhancement","authors":"Kui Jiang;Qiong Wang;Zhaoyi An;Zheng Wang;Cong Zhang;Chia-Wen Lin","doi":"10.1109/TETCI.2024.3369321","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3369321","url":null,"abstract":"Images captured in low-light or underwater environments are often accompanied by significant degradation, which can negatively impact the quality and performance of downstream tasks. While convolutional neural networks (CNNs) and Transformer architectures have made significant progress in computer vision tasks, there are few efforts to harmonize them into a more concise framework for enhancing such images. To this end, this study proposes to aggregate the individual capability of self-attention (SA) and CNNs for accurate perturbation removal while preserving background contents. Based on this, we carry forward a Retinex-based framework, dubbed as Mutual Retinex, where a two-branch structure is designed to characterize the specific knowledge of reflectance and illumination components while removing the perturbation. To maximize its potential, Mutual Retinex is equipped with a new mutual learning mechanism, involving an elaborately designed mutual representation module (MRM). In MRM, the complementary information between reflectance and illumination components are encoded and used to refine each other. Through the complementary learning via the mutual representation, the enhanced results generated by our model exhibit superior color consistency and naturalness. Extensive experiments have shown the significant superiority of our mutual learning based method over thirteen competitors on the low-light task and ten methods on the underwater image enhancement task. In particular, our proposed Mutual Retinex respectively surpasses the state-of-the-art method MIRNet-v2 by 0.90 dB and 2.46 dB in PSNR on the LOL 1000 and FIVEK datasets, while with only 19.8% model parameters.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141096206","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Unmanned aerial vehicles (UAVs) have been widely used in urban missions, and proper planning of UAV paths can improve mission efficiency while reducing the risk of potential third-party impact. Existing work has considered all efficiency and safety objectives for a single decision-maker (DM) and regarded this as a multiobjective optimization problem (MOP). However, there is usually not a single DM but two DMs, i.e., an efficiency DM and a safety DM, and the DMs are only concerned with their respective objectives. The final decision is made based on the solutions of both DMs. In this paper, for the first time, biparty multiobjective UAV path planning (BPMO-UAVPP) problems involving both efficiency and safety departments are modeled. The existing multiobjective immune algorithm with nondominated neighbor-based selection (NNIA), the hybrid evolutionary framework for the multiobjective immune algorithm (HEIA), and the adaptive immune-inspired multiobjective algorithm (AIMA) are modified for solving the BPMO-UAVPP problem, and then biparty multiobjective optimization algorithms, including the BPNNIA, BPHEIA, and BPAIMA, are proposed and comprehensively compared with traditional multiobjective evolutionary algorithms and typical multiparty multiobjective evolutionary algorithms (i.e., OptMPNDS and OptMPNDS2). The experimental results show that BPAIMA performs better than ordinary multiobjective evolutionary algorithms such as NSGA-II and multiparty multiobjective evolutionary algorithms such as OptMPNDS, OptMPNDS2, BPNNIA and BPHEIA.
{"title":"Evolutionary Biparty Multiobjective UAV Path Planning: Problems and Empirical Comparisons","authors":"Kesheng Chen;Wenjian Luo;Xin Lin;Zhen Song;Yatong Chang","doi":"10.1109/TETCI.2024.3361755","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3361755","url":null,"abstract":"Unmanned aerial vehicles (UAVs) have been widely used in urban missions, and proper planning of UAV paths can improve mission efficiency while reducing the risk of potential third-party impact. Existing work has considered all efficiency and safety objectives for a single decision-maker (DM) and regarded this as a multiobjective optimization problem (MOP). However, there is usually not a single DM but two DMs, i.e., an efficiency DM and a safety DM, and the DMs are only concerned with their respective objectives. The final decision is made based on the solutions of both DMs. In this paper, for the first time, biparty multiobjective UAV path planning (BPMO-UAVPP) problems involving both efficiency and safety departments are modeled. The existing multiobjective immune algorithm with nondominated neighbor-based selection (NNIA), the hybrid evolutionary framework for the multiobjective immune algorithm (HEIA), and the adaptive immune-inspired multiobjective algorithm (AIMA) are modified for solving the BPMO-UAVPP problem, and then biparty multiobjective optimization algorithms, including the BPNNIA, BPHEIA, and BPAIMA, are proposed and comprehensively compared with traditional multiobjective evolutionary algorithms and typical multiparty multiobjective evolutionary algorithms (i.e., OptMPNDS and OptMPNDS2). The experimental results show that BPAIMA performs better than ordinary multiobjective evolutionary algorithms such as NSGA-II and multiparty multiobjective evolutionary algorithms such as OptMPNDS, OptMPNDS2, BPNNIA and BPHEIA.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141096323","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","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/TETCI.2024.3369976
Guangliang He;Zhen Zhang;Hanrui Wu;Sanchuan Luo;Yudong Liu
Knowledge graph (KG) is increasingly important in improving recommendation performance and handling item cold-start. A recent research hotspot is designing end-to-end models based on information propagation schemes. However, existing these methods do not highlight key collaborative signals hidden in user-item bipartite graphs, which leads to two problems: (1) the collaborative signal of user collaborative neighbors is not modeled and (2) the incompleteness of KG and the behavioral similarity of item collaborative neighbors are not considered. In this paper, we design a new model called Knowledge Graph Collaborative Neighbor Awareness network (KGCNA) in order to resolve the above problems. KGCNA models the top-k collaborative neighbors of users and items to extract the collaborative preference of the user's top-k collaborative neighbors, the missing attributes of items, and the behavioral similarity of the item's top-k collaborative neighbors, respectively. At the same time, KGCNA designs a novel information aggregation method, which adopts different aggregation methods for users and items to capture the user's item-based behavior preference and the item's long-distance knowledge association in KG, respectively. Furthermore, KGCNA uses an information-gated aggregation mechanism to extract discriminative signals to better study user behavior intent. Experimental results on three benchmark datasets demonstrate that KGCNA significantly improves over state-of-the-art techniques such as CKAN, KGIN, and KGAT.
知识图谱（KG）在提高推荐性能和处理项目冷启动方面越来越重要。最近的一个研究热点是设计基于信息传播方案的端到端模型。然而，现有的这些方法并没有突出隐藏在用户-物品双向图中的关键协作信号，这导致了两个问题：（1）用户协作邻居的协作信号没有被建模；（2）KG 的不完整性和物品协作邻居的行为相似性没有被考虑。为了解决上述问题，我们在本文中设计了一种名为 "知识图谱协作邻居感知网络（KGCNA）"的新模型。KGCNA 对用户和物品的前 k 个协作邻居进行建模，分别提取用户的前 k 个协作邻居的协作偏好、物品的缺失属性和物品的前 k 个协作邻居的行为相似性。同时，KGCNA 设计了一种新颖的信息聚合方法，对用户和物品采用不同的聚合方法，分别捕捉用户基于物品的行为偏好和物品在 KG 中的远距离知识关联。此外，KGCNA 还采用了信息导向聚合机制来提取鉴别信号，从而更好地研究用户行为意图。在三个基准数据集上的实验结果表明，与 CKAN、KGIN 和 KGAT 等最先进的技术相比，KGCNA 的性能有了显著提高。
{"title":"KGCNA: Knowledge Graph Collaborative Neighbor Awareness Network for Recommendation","authors":"Guangliang He;Zhen Zhang;Hanrui Wu;Sanchuan Luo;Yudong Liu","doi":"10.1109/TETCI.2024.3369976","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3369976","url":null,"abstract":"Knowledge graph (KG) is increasingly important in improving recommendation performance and handling item cold-start. A recent research hotspot is designing end-to-end models based on information propagation schemes. However, existing these methods do not highlight key collaborative signals hidden in user-item bipartite graphs, which leads to two problems: (1) the collaborative signal of user collaborative neighbors is not modeled and (2) the incompleteness of KG and the behavioral similarity of item collaborative neighbors are not considered. In this paper, we design a new model called \u0000<italic>Knowledge Graph Collaborative Neighbor Awareness network</i>\u0000 (KGCNA) in order to resolve the above problems. KGCNA models the top-k collaborative neighbors of users and items to extract the collaborative preference of the user's top-k collaborative neighbors, the missing attributes of items, and the behavioral similarity of the item's top-k collaborative neighbors, respectively. At the same time, KGCNA designs a novel information aggregation method, which adopts different aggregation methods for users and items to capture the user's item-based behavior preference and the item's long-distance knowledge association in KG, respectively. Furthermore, KGCNA uses an information-gated aggregation mechanism to extract discriminative signals to better study user behavior intent. Experimental results on three benchmark datasets demonstrate that KGCNA significantly improves over state-of-the-art techniques such as CKAN, KGIN, and KGAT.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141965827","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","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/TETCI.2024.3369636
Xiaoyu Tan;Chao Qu;Junwu Xiong;James Zhang;Xihe Qiu;Yaochu Jin
Model-based reinforcement learning (MBRL) has shown its advantages in sample efficiency over model-free reinforcement learning (MFRL) by leveraging control-based domain knowledge. Despite the impressive results it achieves, MBRL is still outperformed by MFRL due to the lack of unlimited interactions with the environment. While imaginary data can be generated by imagining the trajectories of future states, a trade-off between the usage of data generation and the influence of model bias remains to be resolved. In this paper, we propose a simple and elegant off-policy model-based deep reinforcement learning algorithm with a model embedded in the framework of probabilistic reinforcement learning, called MEMB. To balance the sample-efficiency and model bias, we exploit both real and imaginary data in training. In particular, we embed the model in the policy update and learn value functions from the real data set. We also provide a theoretical analysis of MEMB with the Lipschitz continuity assumption on the model and policy, proving the reliability of the short-term imaginary rollout. Finally, we evaluate MEMB on several benchmarks and demonstrate that our algorithm can achieve state-of-the-art performance.
{"title":"Model-Based Off-Policy Deep Reinforcement Learning With Model-Embedding","authors":"Xiaoyu Tan;Chao Qu;Junwu Xiong;James Zhang;Xihe Qiu;Yaochu Jin","doi":"10.1109/TETCI.2024.3369636","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3369636","url":null,"abstract":"Model-based reinforcement learning (MBRL) has shown its advantages in sample efficiency over model-free reinforcement learning (MFRL) by leveraging control-based domain knowledge. Despite the impressive results it achieves, MBRL is still outperformed by MFRL due to the lack of unlimited interactions with the environment. While imaginary data can be generated by imagining the trajectories of future states, a trade-off between the usage of data generation and the influence of model bias remains to be resolved. In this paper, we propose a simple and elegant off-policy model-based deep reinforcement learning algorithm with a model embedded in the framework of probabilistic reinforcement learning, called MEMB. To balance the sample-efficiency and model bias, we exploit both real and imaginary data in training. In particular, we embed the model in the policy update and learn value functions from the real data set. We also provide a theoretical analysis of MEMB with the Lipschitz continuity assumption on the model and policy, proving the reliability of the short-term imaginary rollout. Finally, we evaluate MEMB on several benchmarks and demonstrate that our algorithm can achieve state-of-the-art performance.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141964720","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","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/TETCI.2024.3369407
Baligh Al-Helali;Qi Chen;Bing Xue;Mengjie Zhang
Symbolic regression is increasingly important for discovering mathematical models for various prediction tasks. It works by searching for the arithmetic expressions that best represent a target variable using a set of input features. However, as the number of features increases, the search process becomes more complex. To address high-dimensional symbolic regression, this work proposes a genetic programming for feature selection method based on the impact of feature removal on the performance of SR models. Unlike existing Shapely value methods that simulate feature absence at the data level, the proposed approach suggests removing features at the model level. This approach circumvents the production of unrealistic data instances, which is a major limitation of Shapely value and permutation-based methods. Moreover, after calculating the importance of the features, a cut-off strategy, which works by injecting a number of random features and utilising their importance to automatically set a threshold, is proposed for selecting important features. The experimental results on artificial and real-world high-dimensional data sets show that, compared with state-of-the-art feature selection methods using the permutation importance and Shapely value, the proposed method not only improves the SR accuracy but also selects smaller sets of features.
符号回归对于发现各种预测任务的数学模型越来越重要。它的工作原理是利用一组输入特征，搜索最能代表目标变量的算术表达式。然而，随着特征数量的增加，搜索过程也变得更加复杂。为了解决高维符号回归问题，本研究根据特征去除对 SR 模型性能的影响，提出了一种遗传编程特征选择方法。与现有的在数据层面模拟特征缺失的 Shapely 值方法不同，所提出的方法建议在模型层面去除特征。这种方法避免了产生不切实际的数据实例，而这正是 Shapely 值和基于排列的方法的主要局限。此外，在计算特征的重要性后，还提出了一种截断策略，即通过注入一些随机特征并利用其重要性自动设置阈值，来选择重要特征。在人工和真实世界高维数据集上的实验结果表明，与使用置换重要性和 Shapely 值的最先进特征选择方法相比，所提出的方法不仅提高了 SR 的准确性，而且选择的特征集更小。
{"title":"Genetic Programming for Feature Selection Based on Feature Removal Impact in High-Dimensional Symbolic Regression","authors":"Baligh Al-Helali;Qi Chen;Bing Xue;Mengjie Zhang","doi":"10.1109/TETCI.2024.3369407","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3369407","url":null,"abstract":"Symbolic regression is increasingly important for discovering mathematical models for various prediction tasks. It works by searching for the arithmetic expressions that best represent a target variable using a set of input features. However, as the number of features increases, the search process becomes more complex. To address high-dimensional symbolic regression, this work proposes a genetic programming for feature selection method based on the impact of feature removal on the performance of SR models. Unlike existing Shapely value methods that simulate feature absence at the data level, the proposed approach suggests removing features at the model level. This approach circumvents the production of unrealistic data instances, which is a major limitation of Shapely value and permutation-based methods. Moreover, after calculating the importance of the features, a cut-off strategy, which works by injecting a number of random features and utilising their importance to automatically set a threshold, is proposed for selecting important features. The experimental results on artificial and real-world high-dimensional data sets show that, compared with state-of-the-art feature selection methods using the permutation importance and Shapely value, the proposed method not only improves the SR accuracy but also selects smaller sets of features.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141096223","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","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/TETCI.2024.3369981
Mehmet Önder Efe;Burak Kürkçü;Coşku Kasnakoǧlu;Zaharuddin Mohamed;Zhijie Liu
This paper introduces the notion of switched neural networks for learning multiple functions under different switching configurations. The neural network structure has adjustable parameters and for each function the state of the parameter vector is determined by a mask vector, 1/0 for active/inactive or +1/-1 for plain/inverted. The optimization problem is to schedule the switching strategy (mask vector) required for each function together with the best parameter vector (weights/biases) minimizing the loss function. This requires a procedure that optimizes a vector containing real and binary values simultaneously to discover commonalities among various functions. Our studies show that a small sized neural network structure with an appropriate switching regime is able to learn multiple functions successfully. During the tests focusing on classification, we considered 2-variable binary functions and all 16 combinations have been chosen as the functions. The regression tests consider four functions of two variables. Our studies showed that simple NN structures are capable of storing multiple information via appropriate switching.
本文介绍了在不同开关配置下学习多种功能的开关神经网络概念。神经网络结构具有可调参数，对于每个功能，参数向量的状态由掩码向量决定，1/0 表示主动/不主动，+1/-1 表示普通/反转。优化问题是安排每个功能所需的切换策略（掩码向量），以及使损失函数最小化的最佳参数向量（权重/偏置）。这就需要同时优化包含实值和二进制值的向量，以发现各种功能之间的共性。我们的研究表明，采用适当切换机制的小型神经网络结构能够成功学习多种函数。在以分类为重点的测试中，我们考虑了双变量二元函数，并选择了所有 16 种组合作为函数。回归测试考虑了两个变量的四个函数。我们的研究表明，简单的 NN 结构能够通过适当的切换存储多种信息。
{"title":"Switched Neural Networks for Simultaneous Learning of Multiple Functions","authors":"Mehmet Önder Efe;Burak Kürkçü;Coşku Kasnakoǧlu;Zaharuddin Mohamed;Zhijie Liu","doi":"10.1109/TETCI.2024.3369981","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3369981","url":null,"abstract":"This paper introduces the notion of switched neural networks for learning multiple functions under different switching configurations. The neural network structure has adjustable parameters and for each function the state of the parameter vector is determined by a mask vector, 1/0 for active/inactive or +1/-1 for plain/inverted. The optimization problem is to schedule the switching strategy (mask vector) required for each function together with the best parameter vector (weights/biases) minimizing the loss function. This requires a procedure that optimizes a vector containing real and binary values simultaneously to discover commonalities among various functions. Our studies show that a small sized neural network structure with an appropriate switching regime is able to learn multiple functions successfully. During the tests focusing on classification, we considered 2-variable binary functions and all 16 combinations have been chosen as the functions. The regression tests consider four functions of two variables. Our studies showed that simple NN structures are capable of storing multiple information via appropriate switching.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141964885","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
In value-based deep reinforcement learning (RL), value function approximation errors lead to suboptimal policies. Temporal difference (TD) learning is one of the most important methodologies to approximate state-action ( $Q$ ) value function. In TD learning, it is critical to estimate $Q$ values of greedy actions more accurately because a more accurate target $Q$ value enhances the estimation accuracy of $Q$ value. To improve the estimation accuracy of $Q$ value, we propose an action-ranked TD learning method to enhance the performance of deep RL by weighting each TD error according to the rank of its corresponding state-action pair's value among all the $Q$ values on a state. The proposed method can provide more accurate target values for TD learning, making the estimation of the $Q$ value more accurate. We apply the proposed method to a representative value-based deep RL algorithm, and results show that the proposed method outperforms baselines on 31 out of 40 Atari games. Furthermore, we extend the proposed method to multi-agent deep RL. To adaptively determine the hyperparameter in action-ranked TD learning, we propose a meta action-ranked TD learning. A series of experiments quantitatively verify that our methods outperform baselines on Atari games, StarCraft-II, and Grid World environments.
{"title":"Data Efficient Deep Reinforcement Learning With Action-Ranked Temporal Difference Learning","authors":"Qi Liu;Yanjie Li;Yuecheng Liu;Ke Lin;Jianqi Gao;Yunjiang Lou","doi":"10.1109/TETCI.2024.3369641","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3369641","url":null,"abstract":"In value-based deep reinforcement learning (RL), value function approximation errors lead to suboptimal policies. Temporal difference (TD) learning is one of the most important methodologies to approximate state-action (\u0000<inline-formula><tex-math>$Q$</tex-math></inline-formula>\u0000) value function. In TD learning, it is critical to estimate \u0000<inline-formula><tex-math>$Q$</tex-math></inline-formula>\u0000 values of greedy actions more accurately because a more accurate target \u0000<inline-formula><tex-math>$Q$</tex-math></inline-formula>\u0000 value enhances the estimation accuracy of \u0000<inline-formula><tex-math>$Q$</tex-math></inline-formula>\u0000 value. To improve the estimation accuracy of \u0000<inline-formula><tex-math>$Q$</tex-math></inline-formula>\u0000 value, we propose an action-ranked TD learning method to enhance the performance of deep RL by weighting each TD error according to the rank of its corresponding state-action pair's value among all the \u0000<inline-formula><tex-math>$Q$</tex-math></inline-formula>\u0000 values on a state. The proposed method can provide more accurate target values for TD learning, making the estimation of the \u0000<inline-formula><tex-math>$Q$</tex-math></inline-formula>\u0000 value more accurate. We apply the proposed method to a representative value-based deep RL algorithm, and results show that the proposed method outperforms baselines on 31 out of 40 Atari games. Furthermore, we extend the proposed method to multi-agent deep RL. To adaptively determine the hyperparameter in action-ranked TD learning, we propose a meta action-ranked TD learning. A series of experiments quantitatively verify that our methods outperform baselines on Atari games, StarCraft-II, and Grid World environments.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141965138","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","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/TETCI.2024.3369313
Junyu Li;Fei Qi;Xin Sun;Bin Zhang;Xiangmin Xu;Hongmin Cai
Unsupervised feature selection is vital in explanatory learning and remains challenging due to the difficulty of formulating a learnable model. Recently, graph embedding learning has gained widespread popularity in unsupervised learning, which extracts low-dimensional representation based on graph structure. Nevertheless, such an embedding scheme for unsupervised feature selection will distort original features due to the spatial transformation by extraction. To address this problem, this paper proposes a collaborative graph embedding model for unsupervised feature selection via jointly using soft-threshold and low-dimensional embedding learning. The former learns a threshold selection matrix for feature weighting in the original space. The latter extracts embedded representation in low-dimensional space to reveal the latent graph structure. By collaborative learning, the proposed method can simultaneously perform unsupervised feature selection in the original space and adaptive graph learning via dual embedding. Extensive experiments on five benchmark datasets demonstrate that the proposed method achieves superior performance compared to eight competing methods.
{"title":"Unsupervised Feature Selection via Collaborative Embedding Learning","authors":"Junyu Li;Fei Qi;Xin Sun;Bin Zhang;Xiangmin Xu;Hongmin Cai","doi":"10.1109/TETCI.2024.3369313","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3369313","url":null,"abstract":"Unsupervised feature selection is vital in explanatory learning and remains challenging due to the difficulty of formulating a learnable model. Recently, graph embedding learning has gained widespread popularity in unsupervised learning, which extracts low-dimensional representation based on graph structure. Nevertheless, such an embedding scheme for unsupervised feature selection will distort original features due to the spatial transformation by extraction. To address this problem, this paper proposes a collaborative graph embedding model for unsupervised feature selection via jointly using soft-threshold and low-dimensional embedding learning. The former learns a threshold selection matrix for feature weighting in the original space. The latter extracts embedded representation in low-dimensional space to reveal the latent graph structure. By collaborative learning, the proposed method can simultaneously perform unsupervised feature selection in the original space and adaptive graph learning via dual embedding. Extensive experiments on five benchmark datasets demonstrate that the proposed method achieves superior performance compared to eight competing methods.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141096296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-08DOI: 10.1109/TETCI.2024.3369868
Savas Ozkan;M. Alper Selver;Bora Baydar;Ali Emre Kavur;Cemre Candemir;Gozde Bozdagi Akar
Despite the widespread use of deep learning methods for semantic segmentation from single imaging modalities, their performance for exploiting multi-domain data still needs to improve. However, the decision-making process in radiology is often guided by data from multiple sources, such as pre-operative evaluation of living donated liver transplantation donors. In such cases, cross-modality performances of deep models become more important. Unfortunately, the domain-dependency of existing techniques limits their clinical acceptability, primarily confining their performance to individual domains. This issue is further formulated as a multi-source domain adaptation problem, which is an emerging field mainly due to the diverse pattern characteristics exhibited from cross-modality data. This paper presents a novel method that can learn robust representations from unpaired cross-modal (CT-MR) data by encapsulating distinct and shared patterns from multiple modalities. In our solution, the covariate shift property is maintained with structural modifications in our architecture. Also, an adversarial loss is adopted to boost the representation capacity. As a result, sparse and rich representations are obtained. Another superiority of our model is that no information about modalities is needed at the training or inference phase. Tests on unpaired CT and MR liver data obtained from the cross-modality task of the CHAOS grand challenge demonstrate that our approach achieves state-of-the-art results with a large margin in both individual metrics and overall scores.
{"title":"Cross-Modal Learning via Adversarial Loss and Covariate Shift for Enhanced Liver Segmentation","authors":"Savas Ozkan;M. Alper Selver;Bora Baydar;Ali Emre Kavur;Cemre Candemir;Gozde Bozdagi Akar","doi":"10.1109/TETCI.2024.3369868","DOIUrl":"https://doi.org/10.1109/TETCI.2024.3369868","url":null,"abstract":"Despite the widespread use of deep learning methods for semantic segmentation from single imaging modalities, their performance for exploiting multi-domain data still needs to improve. However, the decision-making process in radiology is often guided by data from multiple sources, such as pre-operative evaluation of living donated liver transplantation donors. In such cases, cross-modality performances of deep models become more important. Unfortunately, the domain-dependency of existing techniques limits their clinical acceptability, primarily confining their performance to individual domains. This issue is further formulated as a multi-source domain adaptation problem, which is an emerging field mainly due to the diverse pattern characteristics exhibited from cross-modality data. This paper presents a novel method that can learn robust representations from unpaired cross-modal (CT-MR) data by encapsulating distinct and shared patterns from multiple modalities. In our solution, the covariate shift property is maintained with structural modifications in our architecture. Also, an adversarial loss is adopted to boost the representation capacity. As a result, sparse and rich representations are obtained. Another superiority of our model is that no information about modalities is needed at the training or inference phase. Tests on unpaired CT and MR liver data obtained from the cross-modality task of the CHAOS grand challenge demonstrate that our approach achieves state-of-the-art results with a large margin in both individual metrics and overall scores.","PeriodicalId":13135,"journal":{"name":"IEEE Transactions on Emerging Topics in Computational Intelligence","volume":null,"pages":null},"PeriodicalIF":5.3,"publicationDate":"2024-03-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141965826","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","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：