Pub Date : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004157
Qingyan Yin, Wangwang Chen, Ruiping Wu, Zhi Wei
Survival analysis with genomics data provides a deep understanding of biological processes related to prognosis and disease progression at the molecular level. However, high-dimensional small sample genome data causes computational challenges in survival analysis. To address this problem of overfitting and poor interpretation of existing models, we applied the deep learning technology to genome data and proposed a survival analysis model based on an image-based residual neural network model, called Cox-ResNet. High-dimensional gene expression data was embedded into 2D images according to gene positions on chromosomes, and then a residual network model based on Cox proportional hazards was introduced to perform survival analysis. We demonstrated the performance of Cox-ResNet on five datasets of different cancer types from TCGA, comparing it with the cutting-edge survival analysis methods. The Cox-ResNet model not only shows better performance in prediction accuracy, but also biologically interpretable, by generating heat-maps and prognostic genes for high-risk groups with the guided Grad-Cam visualization method. By performing protein-protein interaction network analysis, we examined hub genes and their biological functions for the bladder cancer. These findings confirm that Cox-ResNet model provides a new solution for discovering the driver genes of poor cancer prognosis.
{"title":"Cox-ResNet: A Survival Analysis Model Based on Residual Neural Networks for Gene Expression Data","authors":"Qingyan Yin, Wangwang Chen, Ruiping Wu, Zhi Wei","doi":"10.1109/ICNSC55942.2022.10004157","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004157","url":null,"abstract":"Survival analysis with genomics data provides a deep understanding of biological processes related to prognosis and disease progression at the molecular level. However, high-dimensional small sample genome data causes computational challenges in survival analysis. To address this problem of overfitting and poor interpretation of existing models, we applied the deep learning technology to genome data and proposed a survival analysis model based on an image-based residual neural network model, called Cox-ResNet. High-dimensional gene expression data was embedded into 2D images according to gene positions on chromosomes, and then a residual network model based on Cox proportional hazards was introduced to perform survival analysis. We demonstrated the performance of Cox-ResNet on five datasets of different cancer types from TCGA, comparing it with the cutting-edge survival analysis methods. The Cox-ResNet model not only shows better performance in prediction accuracy, but also biologically interpretable, by generating heat-maps and prognostic genes for high-risk groups with the guided Grad-Cam visualization method. By performing protein-protein interaction network analysis, we examined hub genes and their biological functions for the bladder cancer. These findings confirm that Cox-ResNet model provides a new solution for discovering the driver genes of poor cancer prognosis.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"222 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127516334","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 : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004142
Chaoqian Wang, Lixin Zheng, Shuwan Pan
Due to the development of laser radar, depth camera and other technologies, point cloud data is used in more and more fields. However, compared with two-dimensional image data, the cost of manually labeling point cloud data is higher. This paper present a simple contrastive process to obtain the feature extraction encoder of point cloud data through self-supervised learning, which can provide better support for tasks such as classification and segmentation. We translate a mini batch of date into two crops, the corresponding point clouds data are treated as positive example, and the not corresponding data are treated as negative example. Using InfoNCE as target function to get the unique feature of each data. Comparing to other existing contrastive structure, it performs a higher accuracy in classification task based on ModelNet40. At the same time, we used rotation, randomly cutting and randomly dropout point to realize data augmentation based on ModelNet40 for improving the performance of feature extraction.
{"title":"Research On Feature Extraction of Point Cloud Data Based on Contrastive Learning","authors":"Chaoqian Wang, Lixin Zheng, Shuwan Pan","doi":"10.1109/ICNSC55942.2022.10004142","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004142","url":null,"abstract":"Due to the development of laser radar, depth camera and other technologies, point cloud data is used in more and more fields. However, compared with two-dimensional image data, the cost of manually labeling point cloud data is higher. This paper present a simple contrastive process to obtain the feature extraction encoder of point cloud data through self-supervised learning, which can provide better support for tasks such as classification and segmentation. We translate a mini batch of date into two crops, the corresponding point clouds data are treated as positive example, and the not corresponding data are treated as negative example. Using InfoNCE as target function to get the unique feature of each data. Comparing to other existing contrastive structure, it performs a higher accuracy in classification task based on ModelNet40. At the same time, we used rotation, randomly cutting and randomly dropout point to realize data augmentation based on ModelNet40 for improving the performance of feature extraction.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"17 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126133246","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 : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004192
Tingting Leng, Jufeng Wang, Chunfeng Liu
This paper studies the scheduling problem of single-arm cluster tools that mixedly process two different kinds of wafers without sharing and revisiting processing modules (PMs). We balance internal workloads by adjusting the number of PMs used to process wafers, and balance the external workloads by configuring virtual PMs. We derive the scheduling conditions for single-arm cluster tools, which are more relaxed than the existing ones. We can also use less PMs to get the same production cycle time as the existing literature using configuration of virtual PMs only. We give some examples to show the application and power of the theory.
{"title":"Scheduling of single-arm cluster tools mixedly processing two kinds of wafers","authors":"Tingting Leng, Jufeng Wang, Chunfeng Liu","doi":"10.1109/ICNSC55942.2022.10004192","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004192","url":null,"abstract":"This paper studies the scheduling problem of single-arm cluster tools that mixedly process two different kinds of wafers without sharing and revisiting processing modules (PMs). We balance internal workloads by adjusting the number of PMs used to process wafers, and balance the external workloads by configuring virtual PMs. We derive the scheduling conditions for single-arm cluster tools, which are more relaxed than the existing ones. We can also use less PMs to get the same production cycle time as the existing literature using configuration of virtual PMs only. We give some examples to show the application and power of the theory.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"102 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122931681","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 : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004118
Zemiao Peng, Hao Wu
A non-negative latent factorization of tensors (NLFT) model can well model the temporal pattern hidden in non-negative quality-of-service (QoS) data for predicting the unobserved ones with high accuracy. However, existing NLFT models' objective function is based on Euclidean distance, which is only a special case of $beta$-divergence. Hence, can we build a generalized NLFT model via adopting $beta$-divergence to achieve prediction accuracy gain? To tackle this issue, this paper proposes a $beta$-divergence-based NLFT model ($beta$-NLFT). Its ideas are two-fold: 1) building a learning objective with $beta$-divergence to achieve higher prediction accuracy; and 2) implementing self-adaptation of hyper-parameters to improve practicability. Experimental results generated from two dynamic QoS datasets show that the proposed $beta$-NLFT model can achieve the higher prediction accuracy than state-of-the-art models several when predicting the unobserved QoS data.
{"title":"Non-Negative Latent Factorization of Tensors Model Based on $beta$-Divergence for Time-Aware QoS Prediction","authors":"Zemiao Peng, Hao Wu","doi":"10.1109/ICNSC55942.2022.10004118","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004118","url":null,"abstract":"A non-negative latent factorization of tensors (NLFT) model can well model the temporal pattern hidden in non-negative quality-of-service (QoS) data for predicting the unobserved ones with high accuracy. However, existing NLFT models' objective function is based on Euclidean distance, which is only a special case of $beta$-divergence. Hence, can we build a generalized NLFT model via adopting $beta$-divergence to achieve prediction accuracy gain? To tackle this issue, this paper proposes a $beta$-divergence-based NLFT model ($beta$-NLFT). Its ideas are two-fold: 1) building a learning objective with $beta$-divergence to achieve higher prediction accuracy; and 2) implementing self-adaptation of hyper-parameters to improve practicability. Experimental results generated from two dynamic QoS datasets show that the proposed $beta$-NLFT model can achieve the higher prediction accuracy than state-of-the-art models several when predicting the unobserved QoS data.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"148 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"132750903","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 : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004117
Chuan Ma, Jianhong Ye, LuLu Shuai
Predictive process monitoring belongs to one of the branches of process mining, which aims to provide information in order to proactively mitigate risks and losses. In this paper, we investigate outcome-oriented predictive process monitoring and propose a new way of sequence encoding. The approach uses Hidden Markov Models to capture the relationship between sequences and outcomes to be added to feature encoding. This method combines Hidden Markov Models with existing methods to reduce the dimensionality of the feature vectors while maintaining effective accuracy. We choose the index-based encoding and the last-state encoding as baseline, while three machine learning algorithms are selected for the experiments. The experiment results proved that our method has effective results.
{"title":"Research on Last State Based Hidden Markov Models Encoding Algorithm","authors":"Chuan Ma, Jianhong Ye, LuLu Shuai","doi":"10.1109/ICNSC55942.2022.10004117","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004117","url":null,"abstract":"Predictive process monitoring belongs to one of the branches of process mining, which aims to provide information in order to proactively mitigate risks and losses. In this paper, we investigate outcome-oriented predictive process monitoring and propose a new way of sequence encoding. The approach uses Hidden Markov Models to capture the relationship between sequences and outcomes to be added to feature encoding. This method combines Hidden Markov Models with existing methods to reduce the dimensionality of the feature vectors while maintaining effective accuracy. We choose the index-based encoding and the last-state encoding as baseline, while three machine learning algorithms are selected for the experiments. The experiment results proved that our method has effective results.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"2 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"134049068","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 : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004183
Xiang-En Bai, Jing An, Zihong Yu, Hanqiu Bao, Ke-Fan Wang
As an important research topic in graph learning, the performance of node classification has been improved with the development of some new methods, among which graph neural networks (GNNs) have achieved state-of-the-art node classification performance. However, the existing GNN-based methods mainly address the classification problem with balanced distribution of node samples. However, many real application scenarios of graph data usually have a highly skewed class distribution, i.e., the majority classes occupy most of the samples while the minority classes contain only a few samples. When the nodes exhibit an imbalanced class distribution, existing GNN-based methods favor the majority class and under-represent the minority class. Therefore, we propose a novel Kernel Propagation-based model for Imbalanced Node Classification in Graph Convolutional Networks (KINC-GCN). First, we introduce a kernel propagation method as a preprocessing step to exploit higher-order structural features. The node features are enhanced by concatenating the higher-order structural feature matrix with the node feature matrix. Node embeddings are obtained from the enhanced feature and adjacency matrices by a two-layer GCN, and then a self-optimizing cluster analysis and graph reconstruction module are introduced. The self-optimizing cluster analysis module performs cluster analysis on the node embeddings to enhance the representativeness of the node embeddings. The graph reconstruction module uses an inner product decoder to reconstruct the graph structure and minimize the differences between the reconstructed graph and the original graph. The effectiveness of KINC-GCN in node classification is demonstrated by experiments on three real-world imbalanced graph datasets.
{"title":"A Kernel Propagation-Based Graph Convolutional Network Imbalanced Node Classification Model on Graph Data","authors":"Xiang-En Bai, Jing An, Zihong Yu, Hanqiu Bao, Ke-Fan Wang","doi":"10.1109/ICNSC55942.2022.10004183","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004183","url":null,"abstract":"As an important research topic in graph learning, the performance of node classification has been improved with the development of some new methods, among which graph neural networks (GNNs) have achieved state-of-the-art node classification performance. However, the existing GNN-based methods mainly address the classification problem with balanced distribution of node samples. However, many real application scenarios of graph data usually have a highly skewed class distribution, i.e., the majority classes occupy most of the samples while the minority classes contain only a few samples. When the nodes exhibit an imbalanced class distribution, existing GNN-based methods favor the majority class and under-represent the minority class. Therefore, we propose a novel Kernel Propagation-based model for Imbalanced Node Classification in Graph Convolutional Networks (KINC-GCN). First, we introduce a kernel propagation method as a preprocessing step to exploit higher-order structural features. The node features are enhanced by concatenating the higher-order structural feature matrix with the node feature matrix. Node embeddings are obtained from the enhanced feature and adjacency matrices by a two-layer GCN, and then a self-optimizing cluster analysis and graph reconstruction module are introduced. The self-optimizing cluster analysis module performs cluster analysis on the node embeddings to enhance the representativeness of the node embeddings. The graph reconstruction module uses an inner product decoder to reconstruct the graph structure and minimize the differences between the reconstructed graph and the original graph. The effectiveness of KINC-GCN in node classification is demonstrated by experiments on three real-world imbalanced graph datasets.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"28 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122985929","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 : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004101
Hanqiu Bao, Xudong Shi, Jing An, Hao Li, Qi Kang
This paper introduces a generic multi-agent cooperative control framework using finite-time distributed model predictive control, which applies to multiple types and heterogeneous multi-agent systems with directed topology. We design a detect and avoid collisions strategy using the prediction trajectory of agents, and the agents can safely move toward their goals. The formation time upper bound of multi-agents achieves consensus are derived with given connected communication topology. Numerical simulations validated the feasibility and effectiveness of the proposed approach.
{"title":"Generic Multi-agent Cooperative Control via Finite-time Distributed MPC","authors":"Hanqiu Bao, Xudong Shi, Jing An, Hao Li, Qi Kang","doi":"10.1109/ICNSC55942.2022.10004101","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004101","url":null,"abstract":"This paper introduces a generic multi-agent cooperative control framework using finite-time distributed model predictive control, which applies to multiple types and heterogeneous multi-agent systems with directed topology. We design a detect and avoid collisions strategy using the prediction trajectory of agents, and the agents can safely move toward their goals. The formation time upper bound of multi-agents achieves consensus are derived with given connected communication topology. Numerical simulations validated the feasibility and effectiveness of the proposed approach.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"30 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129211610","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 : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004180
Honghao Zhang, Zhongwei Huang, Guangdong Tian
Design for disassembly (DFD) is a basic design technology serving the scrap and recycling stages, which can well relieve the environmental pressure and improve the economic benefits. The interval 2-tuple q-rung orthopair fuzzy sets (I2q-ROFSs) is proposed to better describe the vague of human thinking and avoid information loss/distortion during the information aggregation phase. Then, this paper proposes a weight method for the disassembly technical features of best worst method (BWM), which can get the best weight vector. The 12q-ROFSs is combined with the grey relational analysis (GRA) for the evaluation of the DFD scheme. Finally, conducting a case study and sensitivity analysis demonstrates the effectiveness of the proposed method. Thus, an effective DFD evaluation method can effectively solve the disassembly design scheme selection problem.
{"title":"A novel MADM method of interval 2-tuple q-rung orthopair fuzzy sets and GRA for DFD schemes","authors":"Honghao Zhang, Zhongwei Huang, Guangdong Tian","doi":"10.1109/ICNSC55942.2022.10004180","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004180","url":null,"abstract":"Design for disassembly (DFD) is a basic design technology serving the scrap and recycling stages, which can well relieve the environmental pressure and improve the economic benefits. The interval 2-tuple q-rung orthopair fuzzy sets (I2q-ROFSs) is proposed to better describe the vague of human thinking and avoid information loss/distortion during the information aggregation phase. Then, this paper proposes a weight method for the disassembly technical features of best worst method (BWM), which can get the best weight vector. The 12q-ROFSs is combined with the grey relational analysis (GRA) for the evaluation of the DFD scheme. Finally, conducting a case study and sensitivity analysis demonstrates the effectiveness of the proposed method. Thus, an effective DFD evaluation method can effectively solve the disassembly design scheme selection problem.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"49 3 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115858686","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 : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004095
LuLu Shuai, Jianhong Ye, Chuan Ma
This paper presents an improved harmony search algorithm (OP-HSA) for missing fault data processing, to find a relatively optimal missing data imputation method from the alternatives. In real life, data is usually missing due to equipment failure or staff negligence, and researchers will preprocess the missing data before using it for work. There are more than one hundred methods to impute missing data, it is not easy to find the right imputation method for missing data. Selecting the right data imputation method can get twice the result with half the effort. To solve this challenge, this paper proposes OP-HSA. In the experiment of this paper, the data processed by different imputation methods are fitted with the original dataset and the similarity error of the dataset is compared. The results show that the improved harmony search algorithm can find a more appropriate data imputation method.
{"title":"Missing Fault Data Processing Method Based On Improved Harmony Search Algorithm","authors":"LuLu Shuai, Jianhong Ye, Chuan Ma","doi":"10.1109/ICNSC55942.2022.10004095","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004095","url":null,"abstract":"This paper presents an improved harmony search algorithm (OP-HSA) for missing fault data processing, to find a relatively optimal missing data imputation method from the alternatives. In real life, data is usually missing due to equipment failure or staff negligence, and researchers will preprocess the missing data before using it for work. There are more than one hundred methods to impute missing data, it is not easy to find the right imputation method for missing data. Selecting the right data imputation method can get twice the result with half the effort. To solve this challenge, this paper proposes OP-HSA. In the experiment of this paper, the data processed by different imputation methods are fitted with the original dataset and the similarity error of the dataset is compared. The results show that the improved harmony search algorithm can find a more appropriate data imputation method.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"47 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116225355","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 : 2022-12-15DOI: 10.1109/ICNSC55942.2022.10004049
Lingyun Wang, Yin Liu, Yunshen Zhou
In real-world scenarios, datasets often perform a long-tailed distribution, making it difficult to train neural net-work models that achieve high accuracy across all classes. In this paper, we explore self-supervised learning for the purpose of learning generalized features and propose a score fusion module to integrate outputs from multiple expert models to obtain a unified prediction. Specifically, we take inspiration from the observation that networks trained on a less unbalanced subset of the distribution tend to produce better performance than networks trained on the entire dataset. However, subsets from tail classes are not adequately represented due to the limitation of data size, which means that their performance is actually unsatisfactory. Therefore, we employ self-supervised learning (SSL) on the whole dataset to obtain a more generalized and transferable feature representation, resulting in a sufficient improvement in subset performance. Unlike previous work that used knowledge distillation models to distill the models trained on a subset to get a unified student model, we propose a score fusion module that directly exploits and integrates the predictions of the subset models. We do extensive experiments on several long-tailed recognition benchmarks to demonstrate the effectiveness of our pronosed model.
{"title":"SFME: Score Fusion from Multiple Experts for Long-tailed Recognition","authors":"Lingyun Wang, Yin Liu, Yunshen Zhou","doi":"10.1109/ICNSC55942.2022.10004049","DOIUrl":"https://doi.org/10.1109/ICNSC55942.2022.10004049","url":null,"abstract":"In real-world scenarios, datasets often perform a long-tailed distribution, making it difficult to train neural net-work models that achieve high accuracy across all classes. In this paper, we explore self-supervised learning for the purpose of learning generalized features and propose a score fusion module to integrate outputs from multiple expert models to obtain a unified prediction. Specifically, we take inspiration from the observation that networks trained on a less unbalanced subset of the distribution tend to produce better performance than networks trained on the entire dataset. However, subsets from tail classes are not adequately represented due to the limitation of data size, which means that their performance is actually unsatisfactory. Therefore, we employ self-supervised learning (SSL) on the whole dataset to obtain a more generalized and transferable feature representation, resulting in a sufficient improvement in subset performance. Unlike previous work that used knowledge distillation models to distill the models trained on a subset to get a unified student model, we propose a score fusion module that directly exploits and integrates the predictions of the subset models. We do extensive experiments on several long-tailed recognition benchmarks to demonstrate the effectiveness of our pronosed model.","PeriodicalId":230499,"journal":{"name":"2022 IEEE International Conference on Networking, Sensing and Control (ICNSC)","volume":"309 2 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-12-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114284897","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：