Phase retrieval (PR) consists of recovering the phase information from captured intensity measurements, known as coded diffraction patterns (CDPs). Non-convex algorithms for addressing the PR problem require a proper initialization that is refined through a gradient descent approach. These PR algorithms have proven to be robust for different scenarios. Despite deep models showing surprising results in this area, these approaches lack interpretability in their neural architectures. This work proposes unrolling the initialization and iterative reconstruction algorithm for the PR problem using the near-field model based on a non-convex formulation; resulting in an interpretable deep neural network (DNN) that can be trained in an end-to-end (E2E) manner. Furthermore, the proposed method can jointly optimize the phase mask for the CDP acquisition and the DNN parameters. Simulation results demonstrate that the proposed E2E method provides high-quality reconstruction using a learned phase mask from a single projection. Also, the proposed method is tested over an experimental optical setup that incorporates the learned phase mask via an only-phase spatial light modulator.
{"title":"Deep Unrolled Single Snapshot Phase Retrieval via Non-Convex Formulation and Phase Mask Design","authors":"Andrés Jerez;Juan Estupiñán;Jorge Bacca;Henry Arguello","doi":"10.1109/JSTSP.2024.3395979","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3395979","url":null,"abstract":"Phase retrieval (PR) consists of recovering the phase information from captured intensity measurements, known as coded diffraction patterns (CDPs). Non-convex algorithms for addressing the PR problem require a proper initialization that is refined through a gradient descent approach. These PR algorithms have proven to be robust for different scenarios. Despite deep models showing surprising results in this area, these approaches lack interpretability in their neural architectures. This work proposes unrolling the initialization and iterative reconstruction algorithm for the PR problem using the near-field model based on a non-convex formulation; resulting in an interpretable deep neural network (DNN) that can be trained in an end-to-end (E2E) manner. Furthermore, the proposed method can jointly optimize the phase mask for the CDP acquisition and the DNN parameters. Simulation results demonstrate that the proposed E2E method provides high-quality reconstruction using a learned phase mask from a single projection. Also, the proposed method is tested over an experimental optical setup that incorporates the learned phase mask via an only-phase spatial light modulator.","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 4","pages":"694-703"},"PeriodicalIF":8.7,"publicationDate":"2024-03-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142587619","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-06DOI: 10.1109/JSTSP.2024.3394970
Xiangrong Wang;Weitong Zhai;Xianghua Wang;Moeness G. Amin;Kaiquan Cai
With the deployment of extremely large-scale array (XL-array) operating at the high frequency bands in future wireless systems, integrated sensing and communication (ISAC) is expected to function in the electromagnetic near-field region with a potential distance of hundreds of meters. Also, a wide signal bandwidth is employed to benefit both communication capacity and sensing resolution. However, most existing works assume a far-field narrowband model, which has prohibited their practical applications in future ISAC systems. In this article, we propose a near-field wideband ISAC framework for concurrent multi-user downlink communications and multi-target localization. In particular, the expression of Cramer Rao Bound (CRB) of direction-of-arrival (DOA) and distance estimations for sensing multiple wideband sources is derived, which is minimized subject to the guaranteed communication quality of service (QoS) for each user. Based on the proposed ISAC framework, sparse transceiver array and the precoding matrix are jointly optimized to reduce mutual coupling and system overhead. The problem is relaxed to a convex optimization and solved iteratively. Simulation results demonstrate that the proposed wideband near-field ISAC framework can well support both modalities and that the sparse transceiver improves the sensing accuracy without sacrificing the communication performance.
{"title":"Wideband Near-Field Integrated Sensing and Communication With Sparse Transceiver Design","authors":"Xiangrong Wang;Weitong Zhai;Xianghua Wang;Moeness G. Amin;Kaiquan Cai","doi":"10.1109/JSTSP.2024.3394970","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3394970","url":null,"abstract":"With the deployment of extremely large-scale array (XL-array) operating at the high frequency bands in future wireless systems, integrated sensing and communication (ISAC) is expected to function in the electromagnetic near-field region with a potential distance of hundreds of meters. Also, a wide signal bandwidth is employed to benefit both communication capacity and sensing resolution. However, most existing works assume a far-field narrowband model, which has prohibited their practical applications in future ISAC systems. In this article, we propose a near-field wideband ISAC framework for concurrent multi-user downlink communications and multi-target localization. In particular, the expression of Cramer Rao Bound (CRB) of direction-of-arrival (DOA) and distance estimations for sensing multiple wideband sources is derived, which is minimized subject to the guaranteed communication quality of service (QoS) for each user. Based on the proposed ISAC framework, sparse transceiver array and the precoding matrix are jointly optimized to reduce mutual coupling and system overhead. The problem is relaxed to a convex optimization and solved iteratively. Simulation results demonstrate that the proposed wideband near-field ISAC framework can well support both modalities and that the sparse transceiver improves the sensing accuracy without sacrificing the communication performance.","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 4","pages":"662-677"},"PeriodicalIF":8.7,"publicationDate":"2024-03-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142587623","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-01DOI: 10.1109/JSTSP.2024.3387274
Shashank Jere;Lizhong Zheng;Karim Said;Lingjia Liu
Recurrent neural networks (RNNs) are known to be universal approximators of dynamic systems under fairly mild and general assumptions. However, RNNs usually suffer from the issues of vanishing and exploding gradients in standard RNN training. Reservoir computing (RC), a special RNN where the recurrent weights are randomized and left untrained, has been introduced to overcome these issues and has demonstrated superior empirical performance especially in scenarios where training samples are extremely limited. On the other hand, the theoretical grounding to support this observed performance has yet been fully developed. In this article, we show that RC can universally approximate a general linear time-invariant (LTI) system. Specifically, we present a clear signal processing interpretation of RC and utilize this understanding in the problem of approximating a generic LTI system. Under this setup, we analytically characterize the optimum probability density function for configuring (instead of training and/or randomly generating) the recurrent weights of the underlying RNN of the RC. Extensive numerical evaluations are provided to validate the optimality of the derived distribution for configuring the recurrent weights of the RC to approximate a general LTI system. Our work results in clear signal processing-based model interpretability of RC and provides theoretical explanation/justification for the power of randomness in randomly generating instead of training RC's recurrent weights. Furthermore, it provides a complete optimum analytical characterization for configuring the untrained recurrent weights, marking an important step towards explainable machine learning (XML) to incorporate domain knowledge for efficient learning.
{"title":"Universal Approximation of Linear Time-Invariant (LTI) Systems Through RNNs: Power of Randomness in Reservoir Computing","authors":"Shashank Jere;Lizhong Zheng;Karim Said;Lingjia Liu","doi":"10.1109/JSTSP.2024.3387274","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3387274","url":null,"abstract":"Recurrent neural networks (RNNs) are known to be universal approximators of dynamic systems under fairly mild and general assumptions. However, RNNs usually suffer from the issues of vanishing and exploding gradients in standard RNN training. Reservoir computing (RC), a special RNN where the recurrent weights are randomized and left untrained, has been introduced to overcome these issues and has demonstrated superior empirical performance especially in scenarios where training samples are extremely limited. On the other hand, the theoretical grounding to support this observed performance has yet been fully developed. In this article, we show that RC can universally approximate a general linear time-invariant (LTI) system. Specifically, we present a clear signal processing interpretation of RC and utilize this understanding in the problem of approximating a generic LTI system. Under this setup, we analytically characterize the optimum probability density function for configuring (instead of training and/or randomly generating) the recurrent weights of the underlying RNN of the RC. Extensive numerical evaluations are provided to validate the optimality of the derived distribution for configuring the recurrent weights of the RC to approximate a general LTI system. Our work results in clear signal processing-based model interpretability of RC and provides theoretical explanation/justification for the power of randomness in randomly generating instead of training RC's recurrent weights. Furthermore, it provides a complete optimum analytical characterization for configuring the untrained recurrent weights, marking an important step towards explainable machine learning (XML) to incorporate domain knowledge for efficient learning.","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 2","pages":"184-198"},"PeriodicalIF":8.7,"publicationDate":"2024-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141500319","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-01DOI: 10.1109/JSTSP.2024.3407390
{"title":"IEEE Signal Processing Society Information","authors":"","doi":"10.1109/JSTSP.2024.3407390","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3407390","url":null,"abstract":"","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 2","pages":"C2-C2"},"PeriodicalIF":8.7,"publicationDate":"2024-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10584141","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141500346","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-01DOI: 10.1109/JSTSP.2024.3398833
Lucas N. Ribeiro;Stefan Schwarz;Markus Rupp;André L. F. de Almeida
Presents corrections to the paper, Energy Efficiency of Mmwave Massive MIMO Precoding With Low-Resolution DACs.
介绍对论文《使用低分辨率 DAC 的 Mmwave Massive MIMO Precoding 的能效》的更正。
{"title":"Corrections to “Energy Efficiency of Mmwave Massive MIMO Precoding With Low-Resolution DACs”","authors":"Lucas N. Ribeiro;Stefan Schwarz;Markus Rupp;André L. F. de Almeida","doi":"10.1109/JSTSP.2024.3398833","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3398833","url":null,"abstract":"Presents corrections to the paper, Energy Efficiency of Mmwave Massive MIMO Precoding With Low-Resolution DACs.","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 2","pages":"259-260"},"PeriodicalIF":8.7,"publicationDate":"2024-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10584340","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141500345","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-01DOI: 10.1109/JSTSP.2024.3417111
Xiao-Ping Zhang
Machine learning methods are the backbone of AI and data science. As we embrace the era of deep learning and foundation models, marked by improving performance due to enhanced computing hardware and data scaling, numerous challenges remain. In areas such as healthcare, government decision making and scientific fields, there is a pressing need to develop transparent models that generate interpretable results. Equally important is the reliability of these models, which ensures their robustness and their ability to generalize to new datasets. This is particularly crucial in sectors like healthcare and autonomous driving, where the stakes are high. The recent trend of increasing model sizes also raises concerns about the environmental and societal impacts of training and deploying large models. Therefore, it is essential to mitigate these impacts as much as possible.
{"title":"Editorial Introduction for the Special Series (Part I) on AI in Signal & Data Science - Toward Explainable, Reliable, and Sustainable Machine Learning","authors":"Xiao-Ping Zhang","doi":"10.1109/JSTSP.2024.3417111","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3417111","url":null,"abstract":"Machine learning methods are the backbone of AI and data science. As we embrace the era of deep learning and foundation models, marked by improving performance due to enhanced computing hardware and data scaling, numerous challenges remain. In areas such as healthcare, government decision making and scientific fields, there is a pressing need to develop transparent models that generate interpretable results. Equally important is the reliability of these models, which ensures their robustness and their ability to generalize to new datasets. This is particularly crucial in sectors like healthcare and autonomous driving, where the stakes are high. The recent trend of increasing model sizes also raises concerns about the environmental and societal impacts of training and deploying large models. Therefore, it is essential to mitigate these impacts as much as possible.","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 2","pages":"139-141"},"PeriodicalIF":8.7,"publicationDate":"2024-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10584285","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141500317","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-01DOI: 10.1109/JSTSP.2024.3407394
{"title":"IEEE Signal Processing Society Information","authors":"","doi":"10.1109/JSTSP.2024.3407394","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3407394","url":null,"abstract":"","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 2","pages":"C3-C3"},"PeriodicalIF":8.7,"publicationDate":"2024-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10584341","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141500318","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-03-01DOI: 10.1109/JSTSP.2024.3381335
Kunal Pattanayak;Vikram Krishnamurthy;Adit Jain
Can deep convolutional neural networks (CNNs) for image classification be interpreted as utility maximizers with information costs? By performing set-valued system identification for Bayesian decision systems, we demonstrate that deep CNNs behave equivalently (in terms of necessary and sufficient conditions) to rationally inattentive Bayesian utility maximizers, a generative model used extensively in economics for human decision-making. Our claim is based on approximately 500 numerical experiments on 5 widely used neural network architectures. The parameters of the resulting interpretable model are computed efficiently via convex feasibility algorithms. As a practical application, we also illustrate how the reconstructed interpretable model can predict the classification performance of deep CNNs with high accuracy. The theoretical foundation of our approach lies in Bayesian revealed preference studied in micro-economics. All our results are on GitHub and completely reproducible.
{"title":"Interpretable Deep Image Classification Using Rationally Inattentive Utility Maximization","authors":"Kunal Pattanayak;Vikram Krishnamurthy;Adit Jain","doi":"10.1109/JSTSP.2024.3381335","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3381335","url":null,"abstract":"Can deep convolutional neural networks (CNNs) for image classification be interpreted as utility maximizers with information costs? By performing set-valued system identification for Bayesian decision systems, we demonstrate that deep CNNs behave equivalently (in terms of necessary and sufficient conditions) to rationally inattentive Bayesian utility maximizers, a generative model used extensively in economics for human decision-making. Our claim is based on approximately 500 numerical experiments on 5 widely used neural network architectures. The parameters of the resulting interpretable model are computed efficiently via convex feasibility algorithms. As a practical application, we also illustrate how the reconstructed interpretable model can predict the classification performance of deep CNNs with high accuracy. The theoretical foundation of our approach lies in Bayesian revealed preference studied in micro-economics. All our results are on GitHub and completely reproducible.","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 2","pages":"168-183"},"PeriodicalIF":8.7,"publicationDate":"2024-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141500320","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-02-22DOI: 10.1109/JSTSP.2024.3368661
Sicong Liu;Xianyao Wang;Jian Song;Zhu Han
This article presents a compressed sensing (CS) based framework for visible light positioning (VLP), designed to achieve simultaneous and precise localization of multiple intelligent robots within an indoor factory. The framework leverages light-emitting diodes (LEDs) originally intended for illumination purposes as anchors, repurposing them for the localization of robots equipped with photodetectors. By predividing the plane encompassing the robot positions into a grid, with the number of robots being notably fewer than the grid points, the inherent sparsity of the arrangement is harnessed. To construct an effective sparse measurement model, a sequence of aggregation, autocorrelation, and cross-correlation operations are employed to the signals. Consequently, the complex task of localizing multiple targets is reformulated into a sparse recovery problem, amenable to resolution through CS-based algorithms. Notably, the localization precision is augmented by inter-target cooperation among the robots, and inter-anchor cooperation among distinct LEDs. Furthermore, to fortify the robustness of localization, a generative adversarial network (GAN) is introduced into the proposed localization framework. The simulation results affirm that the proposed framework can successfully achieve centimeter-level accuracy for simultaneous localization of multiple targets.
本文介绍了一种基于压缩传感（CS）的可见光定位（VLP）框架，旨在实现室内工厂内多个智能机器人的同时精确定位。该框架利用原本用于照明目的的发光二极管（LED）作为锚点，将其重新用于配备光电探测器的机器人的定位。通过将包含机器人位置的平面预先划分为网格，机器人的数量明显少于网格点，从而利用了布置的固有稀疏性。为了构建有效的稀疏测量模型，需要对信号进行一系列聚合、自相关和交叉相关操作。因此，定位多个目标的复杂任务被重新表述为一个稀疏恢复问题，可通过基于 CS 的算法加以解决。值得注意的是，机器人之间的目标间合作以及不同 LED 之间的锚点间合作提高了定位精度。此外，为了加强定位的鲁棒性，在拟议的定位框架中引入了生成对抗网络（GAN）。仿真结果表明，所提出的框架可成功实现厘米级精度的多目标同时定位。
{"title":"Cooperative Robotics Visible Light Positioning: An Intelligent Compressed Sensing and GAN-Enabled Framework","authors":"Sicong Liu;Xianyao Wang;Jian Song;Zhu Han","doi":"10.1109/JSTSP.2024.3368661","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3368661","url":null,"abstract":"This article presents a compressed sensing (CS) based framework for visible light positioning (VLP), designed to achieve simultaneous and precise localization of multiple intelligent robots within an indoor factory. The framework leverages light-emitting diodes (LEDs) originally intended for illumination purposes as anchors, repurposing them for the localization of robots equipped with photodetectors. By predividing the plane encompassing the robot positions into a grid, with the number of robots being notably fewer than the grid points, the inherent sparsity of the arrangement is harnessed. To construct an effective sparse measurement model, a sequence of aggregation, autocorrelation, and cross-correlation operations are employed to the signals. Consequently, the complex task of localizing multiple targets is reformulated into a sparse recovery problem, amenable to resolution through CS-based algorithms. Notably, the localization precision is augmented by inter-target cooperation among the robots, and inter-anchor cooperation among distinct LEDs. Furthermore, to fortify the robustness of localization, a generative adversarial network (GAN) is introduced into the proposed localization framework. The simulation results affirm that the proposed framework can successfully achieve centimeter-level accuracy for simultaneous localization of multiple targets.","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 3","pages":"407-418"},"PeriodicalIF":8.7,"publicationDate":"2024-02-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142137544","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-02-14DOI: 10.1109/JSTSP.2024.3363419
Christopher T. H. Teo;Milad Abdollahzadeh;Ngai-Man Cheung
This work studies fair generative models. We reveal and quantify the biases in state-of-the-art (SOTA) GANs w.r.t. different sensitive attributes. To address the biases, our main contribution is to propose novel methods to learn fair generative models via transfer learning. Specifically, first, we propose FairTL where we pre-train the generative model with a large biased dataset, then adapt the model using a small fair reference dataset. Second, to further improve sample diversity, we propose FairTL++ , containing two additional innovations: 1) aligned feature adaptation , which preserves learned general knowledge while improving fairness by adapting only sensitive attribute-specific parameters, 2) multiple feedback discrimination , which introduces a frozen discriminator for quality feedback and another evolving discriminator for fairness feedback. Taking one step further, we consider an alternative challenging and practical setup. Here, only a pre-trained model is available but the dataset used to pre-train the model is inaccessible. We remark that previous work requires access to large, biased datasets and cannot handle this setup. Extensive experimental results show that FairTL and FairTL++ achieve state-of-the-art performance in quality, diversity and fairness in both setups.
{"title":"FairTL: A Transfer Learning Approach for Bias Mitigation in Deep Generative Models","authors":"Christopher T. H. Teo;Milad Abdollahzadeh;Ngai-Man Cheung","doi":"10.1109/JSTSP.2024.3363419","DOIUrl":"https://doi.org/10.1109/JSTSP.2024.3363419","url":null,"abstract":"This work studies fair generative models. We reveal and quantify the biases in state-of-the-art (SOTA) GANs \u0000<italic>w.r.t.</i>\u0000 different sensitive attributes. To address the biases, our main contribution is to propose novel methods to learn fair generative models via transfer learning. Specifically, first, we propose \u0000<sc>FairTL</small>\u0000 where we pre-train the generative model with a large biased dataset, then adapt the model using a small fair reference dataset. Second, to further improve sample diversity, we propose \u0000<sc>FairTL++</small>\u0000, containing two additional innovations: 1) \u0000<italic>aligned feature adaptation</i>\u0000, which preserves learned general knowledge while improving fairness by adapting only sensitive attribute-specific parameters, 2) \u0000<italic>multiple feedback discrimination</i>\u0000, which introduces a frozen discriminator for quality feedback and another evolving discriminator for fairness feedback. Taking one step further, we consider an alternative challenging and practical setup. Here, only a pre-trained model is available but the dataset used to pre-train the model is inaccessible. We remark that previous work requires access to large, biased datasets and cannot handle this setup. Extensive experimental results show that \u0000<sc>FairTL</small>\u0000 and \u0000<sc>FairTL++</small>\u0000 achieve state-of-the-art performance in quality, diversity and fairness in both setups.","PeriodicalId":13038,"journal":{"name":"IEEE Journal of Selected Topics in Signal Processing","volume":"18 2","pages":"155-167"},"PeriodicalIF":8.7,"publicationDate":"2024-02-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"141500323","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","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：