Latent Feature Space Research Articles

Knowledge Distillation from Single-Task Teachers to Multi-Task Student for End-to-End Autonomous Driving

In the domain of end-to-end autonomous driving, conventional sensor fusion techniques exhibit inadequacies, particularly when facing challenging scenarios with numerous dynamic agents. Imitation learning hampers the performance by the expert and encounters issues with out-of-distribution challenges. To overcome these limitations, we propose a transformer-based algorithm designed to fuse diverse representations from RGB-D cameras through knowledge distillation. This approach leverages insights from multi-task teachers to enhance the learning capabilities of single-task students, particularly in a Reinforcement Learning (RL) setting. Our model consists of two primary modules: the perception module, responsible for encoding observation data acquired from RGB-D cameras and performing tasks such as semantic segmentation, semantic depth cloud mapping (SDC), ego vehicle speed estimation, and traffic light state recognition. Subsequently, the control module decodes these features, incorporating additional data, including a rough simulator for static and dynamic environments, to anticipate waypoints within a latent feature space. Vehicular controls (e.g., steering, throttle, and brake) are obtained directly from measurement features and environmental states using the RL agent and are further refined by a PID algorithm that dynamically follows waypoints. The model undergoes rigorous evaluation and comparative analysis on the CARLA simulator across various scenarios, encompassing normal to adversarial conditions. Our code is available at https://github.com/pagand/e2etransfuser/ to facilitate future studies.

From Hope to Safety: Unlearning Biases of Deep Models via Gradient Penalization in Latent Space

Deep Neural Networks are prone to learning spurious correlations embedded in the training data, leading to potentially biased predictions. This poses risks when deploying these models for high-stake decision-making, such as in medical applications. Current methods for post-hoc model correction either require input-level annotations which are only possible for spatially localized biases, or augment the latent feature space, thereby hoping to enforce the right reasons. We present a novel method for model correction on the concept level that explicitly reduces model sensitivity towards biases via gradient penalization. When modeling biases via Concept Activation Vectors, we highlight the importance of choosing robust directions, as traditional regression-based approaches such as Support Vector Machines tend to result in diverging directions. We effectively mitigate biases in controlled and real-world settings on the ISIC, Bone Age, ImageNet and CelebA datasets using VGG, ResNet and EfficientNet architectures. Code and Appendix are available on https://github.com/frederikpahde/rrclarc.

PixRevive: Latent Feature Diffusion Model for Compressed Video Quality Enhancement.

In recent years, the rapid prevalence of high-definition video in Internet of Things (IoT) systems has been directly facilitated by advances in imaging sensor technology. To adapt to limited uplink bandwidth, most media platforms opt to compress videos to bitrate streams for transmission. However, this compression often leads to significant texture loss and artifacts, which severely degrade the Quality of Experience (QoE). We propose a latent feature diffusion model (LFDM) for compressed video quality enhancement, which comprises a compact edge latent feature prior network (ELPN) and a conditional noise prediction network (CNPN). Specifically, we first pre-train ELPNet to construct a latent feature space that captures rich detail information for representing sharpness latent variables. Second, we incorporate these latent variables into the prediction network to iteratively guide the generation direction, thus resolving the problem that the direct application of diffusion models to temporal prediction disrupts inter-frame dependencies, thereby completing the modeling of temporal correlations. Lastly, we innovatively develop a Grouped Domain Fusion module that effectively addresses the challenges of diffusion distortion caused by naive cross-domain information fusion. Comparative experiments on the MFQEv2 benchmark validate our algorithm's superior performance in terms of both objective and subjective metrics. By integrating with codecs and image sensors, our method can provide higher video quality.

AdaClass filter and its application in bearing fault diagnosis

The vibration signals recorded by the sensor reflect the operating state of bearings, and extracting recognizable features effectively from them has become a hot issue in fault diagnosis. Currently, signal processing based filtering methods have emerged as a popular approach for extracting fault-related features. However, conventional filters based on specified assumptions and theoretical models have limited adaptability to multiple types of bearings under different operating conditions, which can significantly impact the diagnostic results. Given this, a data-driven Adaptive Class (AdaClass) filter is proposed to extract the response characteristics of different categories within the latent space. The filter details are obtained by statistically analyzing the mean vectors of samples for each class in the reconstructed feature subspaces. Notably, the latent feature space is mapped by linear operators linear discriminant analysis and class-wise principal component analysis, where the data has a more concise feature representation and a more distinct feature structure. The low-dimensional projection operations enhance the differential information among different categories, and reorganize the internal structure within the same category. Furthermore, a bearing fault diagnosis model is developed based on the AdaClass filter banks, utilizing one-step convolution to improve the efficiency of feature extraction. Experimental results show that the proposed method outperforms the competitors in terms of accuracy, time consumption, and noise resistance, especially for small sample scenarios.

Preserving domain private information via mutual information maximization

Recent advances in unsupervised domain adaptation have shown that mitigating the domain divergence by extracting the domain-invariant features could significantly improve the generalization of a model with respect to a new data domain. However, current methodologies often neglect to retain domain private information, which is the unique information inherent to the unlabeled new domain, compromising generalization. This paper presents a novel method that utilizes mutual information to protect this domain-specific information, ensuring that the latent features of the unlabeled data not only remain domain-invariant but also reflect the unique statistics of the unlabeled domain. We show that simultaneous maximization of mutual information and reduction of domain divergence can effectively preserve domain-private information. We further illustrate that a neural estimator can aptly estimate the mutual information between the unlabeled input space and its latent feature space. Both theoretical analysis and empirical results validate the significance of preserving such unique information of the unlabeled domain for cross-domain generalization. Comparative evaluations reveal our method’s superiority over existing state-of-the-art techniques across multiple benchmark datasets.

High-low level task combination for object detection in foggy weather conditions

For the object detection task in foggy weather conditions, image dehazing network is often used as preprocessing method to get a clear input. However, there is not strictly a strong positive correlation between image dehazing task and object detection task. Moreover, the preprocessing module can increase the inference time of the whole model to a certain extent. To alleviate these problems, we propose a novel High-Low level task combination network (HLNet) based on multitask learning, which can learn both high-level and low-level tasks. Specially, instead of restoring the features to clear pixel-wise feature space like common image dehazing method, we opt to perform a restoration in feature level to mitigate the influence of the Batch Normalization (BN) layer of encoder on dehazing task. HLNet jointly learn dehazing task and detection task in an end-to-end fashion, which ensures that the weather-specific information in latent feature space is suppressed. Moreover, we applied the HLNet framework on three different object detection networks, including RetinaNnet, YOLOv3 and YOLOv5s network, and achieved improvements of 1.7 percent, 2.3 percent, and 1.2 percent in mAP respectively. The experimental results demonstrate the effectiveness and generalization ability of our proposed HLNet framework in real foggy scenarios.

CloudMix: Dual Mixup Consistency for Unpaired Point Cloud Completion.

Due to the unsatisfactory performance of supervised methods on unpaired real-world scans, point cloud completion via cross-domain adaptation has recently drawn growing attention. Nevertheless, previous approaches only focus on alleviating the distribution shift through domain alignment, resulting in massive information loss of real-world domain data. To tackle this issue, we propose a dual mixup-induced consistency regularization to integrate both source and target domain to improve robustness and generalization capability. Specifically, we mix up virtual and real-world shapes in the input and latent feature space respectively, and then regularize the completion network by forcing two kinds of mixed completion predictions to be consistent. To further adapt to each instance within the real-world domain, we design a novel density-aware refiner to utilize local context information to preserve the fine-grained details and remove noise or outliers for coarse completion. Extensive experiments on real-world scans and our synthetic unpaired datasets demonstrate the superiority of our method over existing state-of-the-art approaches.

On the application of domain adaptation for knowledge transfer and damage detection across bridge spans: an experimental case study

Machine Learning offers a substantial support for vibration-based Structural Health Monitoring and automatic damage assessment in real-time. However, labelled health-state data are often limited and difficult to collect, especially under various environmental conditions and damage scenarios. To bridge this gap, Population-Based Structural Health Monitoring is employed to enrich the available dataset by acquiring information from a population of similar structures and to afterwards develop Transfer Learning strategies across different structural systems. In this field, a major challenge is that the development and validation of robust techniques require a large dataset from multiple real structures covering a wide range of health-states. To this end, the present paper firstly illustrates an experimental campaign conducted on a laboratory-scale bridge model to create a comprehensive dataset using different bridge configurations, that are subjected to various environmental conditions and simulated damage scenarios. It follows that the described population certainly provides the basis to investigate TL and gain insights into the field of PBSHM. Working on this case-study, a new application of Domain Adaptation is proposed, with the aim to perform damage identification via label sharing across different spans of a continuously monitored bridge. The extracted natural frequencies are projected into a latent feature space and a machine learning algorithm is then adopted to recognise a certain damage in one span, based on the information previously learned on a different bridge span. This approach is useful if considering that multiple portions could be reasonably affected by similar defects. Therefore, the proposed approach overcomes the limitations of conventional ML-based SHM methods, paving the way to facilitate risk assessment in transportation networks and identify similar anomalies by exchanging span-related health-state information.

RPM 2.0: RF-Based Pose Machines for Multi-Person 3D Pose Estimation

Advanced human sensing technologies based on radio frequency (RF) signals have gained widespread attention in recent years. However, due to the sparsity and incompleteness of RF signals, fine-grained RF-based multi-person 3D pose estimation has progressed more slowly. In this paper, we present RF-based Pose Machine (RPM 2.0) for multi-person 3D pose estimation using RF signals. Specifically, we first develop a lightweight anchor-free detector module to locate and crop regions of interest from horizontal and vertical RF signals. Afterward, we treat the horizontal and vertical millimeter-wave radars as “RF cameras” with different viewing angles and propose a Multi-view Fusion Network to unproject the RF signals into a unified latent feature space, and then calculate the correlation for weighted fusion. Finally, a Spatio-Temporal Attention Network is designed to reconstruct the multi-person 3D skeleton sequences, in which the spatial attention module is proposed to recover invisible body parts using non-local correlations among joints and the temporal attention module refines the 3D pose sequences using temporal coherency learned from frame queries. We evaluate the performance of the proposed RPM 2.0 and state-of-the-art methods on a large-scale dataset with multi-person 3D pose labels and corresponding radar signals. The experimental results show that RPM 2.0 outperforms all of the baseline methods, which locates multi-person 3D key points with an average error of 73 <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mm</i> and generalizes well in new data such as occlusion, low illumination.

Heterogeneous Domain Adaptation for Multistream Classification on Cyber Threat Data

Under a newly introduced setting of multistream classification, two data streams are involved, which are referred to as source and target streams. The source stream continuously generates data instances from a certain domain with labels, while the target stream does the same task without labels from another domain. Existing approaches assume that domains for both data streams are identical, which is not quite true, since data streams from different sources may contain distinct features. Indeed, they may even have different numbers of features. Furthermore, obtaining labels for every instance in a data stream is often expensive and time-consuming. Therefore, it has become an important topic to explore if classes of labeled instances from other related streams are helpful to predict the classes of unlabeled instances in a different stream. Note that domains of source and target streams may have distinct feature spaces and data distributions. Our objective is to predict class labels of data instances in the target stream by using the classifiers trained by the source stream. We propose a framework of multistream classification by using projected data from a common latent feature space, which is embedded from both source and target domains. This framework is also crucial for enterprise system defenders to detect cross-platform attacks, such as Advanced Persistent Threats (APTs). Empirical valuation and analysis on both real-world and synthetic datasets are performed to validate the effectiveness of our proposed algorithm, comparing to state-of-the-art techniques. Experimental results show that our approach significantly outperforms other existing approaches.

Bidirectional visual-tactile cross-modal generation using latent feature space flow model

Inspired by visual-tactile cross-modal bidirectional mapping of the human brain, this paper introduces a novel approach to bidirectional mapping between visual and tactile data, an area not fully explored in the predominantly unidirectional existing studies. First, we adopt separate Variational AutoEncoder (VAE) models for visual and tactile data. Furthermore, we introduce a conditional flow model built on the VAE latent feature space, enabling cross-modal bidirectional mapping between visual and tactile data using one model. The experimental results show that our method achieves excellent performance in terms of the similarity between the generated data and the original data (Structural Similarity Index (SSIM) of visual data: 0.58, SSIM of tactile data: 0.80), the classification accuracy on generated data (visual data: 91.60%, tactile data: 88.05%), and the zero-shot classification accuracy between generated data and language (visual data: 44.49%, tactile data: 45.03%). To the best of our knowledge, the method proposed in this paper is the first one to utilize a single model to achieve bidirectional mapping between visual and tactile data. Our model and code will be made public after the acceptance of the paper.

Adversarially robust neural networks with feature uncertainty learning and label embedding

Deep neural networks (DNNs) are vulnerable to the attacks of adversarial examples, which bring serious security risks to the learning systems. In this paper, we propose a new defense method to improve the adversarial robustness of DNNs based on stochastic neural networks (SNNs), termed as Margin-SNN. The proposed Margin-SNN mainly includes two modules, i.e., feature uncertainty learning module and label embedding module. The first module introduces uncertainty to the latent feature space by giving each sample a distributional representation rather than a fixed point representation, and leverages the advantages of variational information bottleneck method in achieving good intra-class compactness in latent space. The second module develops a label embedding mechanism to take advantage of the semantic information underlying the labels, which maps the labels into the same latent space with the features, in order to capture the similarity between sample and its class centroid, where a penalty term is equipped to elegantly enlarge the margin between different classes for better inter-class separability. Since no adversarial information is introduced, the proposed model can be learned in standard training to improve adversarial robustness, which is much more efficient than adversarial training. Extensive experiments on data sets MNIST, FASHION MNIST, CIFAR10, CIFAR100 and SVHN demonstrate superior defensive ability of the proposed method. Our code is available at https://github.com/humeng24/Margin-SNN.

A locational false data injection attack detection method in smart grid based on adversarial variational autoencoders

Stealthy FDIA (False Data Injection Attack) is a serious cyber threat that can modify state estimation of smart grid through maliciously altering the measurement data, but can’t be detected by traditional bad data detection system in smart grid. There exist two weakpoints for numerous deep neural networks (DNNs) based data-driven schemes against FDIA. First, they mainly focus on detecting the presence of FDIA, but fail to localize the specific bus/nodes affected. Second, their performance is not sufficiently desirable under small attack, i.e., anomalies caused by the attack closely resemble normal data. To address the above issues, this paper proposes an effective locational FDIA detection framework based on data reconstruction, AT-GVAE, which seamlessly integrates the variational autoencoders (VAEs) and generative adversarial network paradigm. Specifically, our contributions are threefold. First, the proposed AT-GVAE framework is novelly composed of two main modules: the generative VAE_G and discriminative VAE_D that both play dual roles: reconstruct data from jointly learning distributions of data and latent feature space, and meanwhile play Minmax game with adversarial way. Second, multiple-layer gated recurrent units (GRUs) are utilized as the basic structure of the encoder and decoders in both VAEs, to characterize the temporal correlations of measurement data sequence. Additionally, the self-attention mechanism is used to enhance the expressive ability of GRU based VAEs. Then, the anomaly score for each busbar in the smart grid is determined by comparing the residual between the observed measurement and the outputs of VAE_G and VAE_D, enabling the localization of FDIA. Finally, thorough experiments on multiple power systems with series data of total 3456 measurements demonstrate that, in terms of multiple typical metrics, including false alarm rate vs. detection probability, AUC-ROC, and AUC-PR, our proposed framework outperforms the state-of-the-art GRU, VAE and adversarial training based FDIA detection schemes.

Training multi-source domain adaptation network by mutual information estimation and minimization

We address the problem of Multi-Source Domain Adaptation (MSDA), which trains a neural network using multiple labeled source datasets and an unlabeled target dataset, and expects the trained network to well classify the unlabeled target data. The main challenge in this problem is that the datasets are generated by relevant but different joint distributions. In this paper, we propose to address this challenge by estimating and minimizing the mutual information in the network latent feature space, which leads to the alignment of the source joint distributions and target joint distribution simultaneously. Here, the estimation of the mutual information is formulated into a convex optimization problem, such that the global optimal solution can be easily found. We conduct experiments on several public datasets, and show that our algorithm statistically outperforms its competitors. Video and code are available at https://github.com/sentaochen/Mutual-Information-Estimation-and-Minimization.

Quality-driven Gaussian mixture variational probabilistic network for soft sensor application in PET/PA6 polymerization process

Variational autoencoder has been widely used to build soft sensor models as a deep unsupervised feature extractor in recent years. However, the vanilla VAE employs a single diagonal Gaussian distribution as the approximate posterior in its latent feature space, which may lead to inferior performance when solving complex non-Gaussian multimode modeling problem. To address the issue, this paper proposes a quality-driven Gaussian mixture variational probabilistic network-based (QGMVPN) soft sensor model. First, Gaussian mixture density with diagonal covariance matrices and Householder transformations are implemented to characterize the distributions of non-Gaussian multimode feature. Then, a quality-driven strategy is designed to extract quality-related features and a novel gating mechanism is employed to enhance the nonlinear representation ability. Last, the trained decoder, coupled with the sampling space, is directly applied to soft sensor modeling owing to the special extra input of the decoder. The effectiveness of QGMVPN-based model is demonstrated on two polymerization datasets.

The emergence of task-relevant representations in a nonlinear decision-making task

This paper describes the relationship between performance in a decision-making task and the emergence of task-relevant representations. Participants learnt two tasks in which the appropriate response depended on multiple relevant stimuli and the underlying stimulus-outcome associations were governed by a latent feature that participants could discover. We divided participants into good and bad performers based on their overall classification rate and computed behavioural accuracy for each feature value. We found that participants with better performance had a better representation of the latent feature space. We then used representation similarity analysis on Electroencephalographic (EEG) data to identify when these representations emerge. We were able to decode task-relevant representations in a time window emerging 700 ms after stimulus presentation, but only for participants with good task performance. Our findings suggest that, in order to make good decisions, it is necessary to create and extract a low-dimensional representation of the task at hand.

Anchor Link Prediction for Privacy Leakage via De-Anonymization in Multiple Social Networks

Anchor link prediction exacerbates the risk of privacy leakage via the de-anonymization of social network data. Embedding-based methods for anchor link prediction are limited by the excessive similarity of the associated nodes in a latent feature space and the variation between latent feature spaces caused by the semantics of different networks. In this paper, we propose a novel method which reduces the impact of semantic discrepancies between different networks in the latent feature space. The proposed method consists of two phases. First, graph embedding focuses on the network structural roles of nodes and increases the distinction between the associated nodes in the embedding space. Second, a federated adversarial learning framework which performs graph embedding on each social network and an adversarial learning model on the server according to the observable anchor links is used to associate independent graph embedding approaches on different social networks. The combination of distinction enhancement and the association of graph embedding approaches alleviates variance between the latent feature spaces caused by the semantics of different social networks. Extensive experiments on real social networks demonstrate that the proposed method significantly outperforms the state-of-the-art methods in terms of both precision and robustness.

Heterogeneous Regularization-Based Tensor Subspace Clustering for Hyperspectral Band Selection.

Band selection (BS) reduces effectively the spectral dimension of a hyperspectral image (HSI) by selecting relatively few representative bands, which allows efficient processing in subsequent tasks. Existing unsupervised BS methods based on subspace clustering are built on matrix-based models, where each band is reshaped as a vector. They encode the correlation of data only in the spectral mode (dimension) and neglect strong correlations between different modes, i.e., spatial modes and spectral mode. Another issue is that the subspace representation of bands is performed in the raw data space, where the dimension is often excessively high, resulting in a less efficient and less robust performance. To address these issues, in this article, we propose a tensor-based subspace clustering model for hyperspectral BS. Our model is developed on the well-known Tucker decomposition. The three factor matrices and a core tensor in our model encode jointly the multimode correlations of HSI, avoiding effectively to destroy the tensor structure and information loss. In addition, we propose well-motivated heterogeneous regularizations (HRs) on the factor matrices by taking into account the important local and global properties of HSI along three dimensions, which facilitates the learning of the intrinsic cluster structure of bands in the low-dimensional subspaces. Instead of learning the correlations of bands in the original domain, a common way for the matrix-based models, our model learns naturally the band correlations in a low-dimensional latent feature space, which is derived by the projections of two factor matrices associated with spatial dimensions, leading to a computationally efficient model. More importantly, the latent feature space is learned in a unified framework. We also develop an efficient algorithm to solve the resulting model. Experimental results on benchmark datasets demonstrate that our model yields improved performance compared to the state-of-the-art.

SemiMAR: Semi-Supervised Learning for CT Metal Artifact Reduction.

Metal artifacts lead to CT imaging quality degradation. With the success of deep learning (DL) in medical imaging, a number of DL-based supervised methods have been developed for metal artifact reduction (MAR). Nonetheless, fully-supervised MAR methods based on simulated data do not perform well on clinical data due to the domain gap. Although this problem can be avoided in an unsupervised way to a certain degree, severe artifacts cannot be well suppressed in clinical practice. Recently, semi-supervised metal artifact reduction (MAR) methods have gained wide attention due to their ability in narrowing the domain gap and improving MAR performance in clinical data. However, these methods typically require large model sizes, posing challenges for optimization. To address this issue, we propose a novel semi-supervised MAR framework. In our framework, only the artifact-free parts are learned, and the artifacts are inferred by subtracting these clean parts from the metal-corrupted CT images. Our approach leverages a single generator to execute all complex transformations, thereby reducing the model's scale and preventing overlap between clean part and artifacts. To recover more tissue details, we distill the knowledge from the advanced dual-domain MAR network into our model in both image domain and latent feature space. The latent space constraint is achieved via contrastive learning. We also evaluate the impact of different generator architectures by investigating several mainstream deep learning-based MAR backbones. Our experiments demonstrate that the proposed method competes favorably with several state-of-the-art semi-supervised MAR techniques in both qualitative and quantitative aspects.

Semi-Supervised Domain Alignment Learning for Single Image Dehazing.

Convolutional neural networks (CNNs) have attracted much research attention and achieved great improvements in single-image dehazing. However, previous learning-based dehazing methods are mainly trained on synthetic data, which greatly degrades their generalization capability on natural hazy images. To address this issue, this article proposes a semi-supervised learning approach for single-image dehazing, where both synthetic and realistic images are leveraged during training. Considering the situation that it is hard to obtain the realistic pairs of hazy and haze-free images, how to utilize the realistic data is not a trivial work. In this article, a domain alignment module is introduced to narrow the distribution distance between synthetic data and realistic hazy images in a latent feature space. Meanwhile, a haze-aware attention module is designed to describe haze densities of different regions in the image, thus adaptively responds for different hazy areas. Furthermore, the dark channel prior is introduced to the framework to improve the quality of the unsupervised learning results by considering the statistical characters of haze-free images. Such a semi-supervised design can significantly address the domain shift issue between the synthetic and realistic data, and improve generalization performance in the real world. Experiments indicate that the proposed method obtains state-of-the-art performance on both public synthetic and realistic hazy images with better visual results.