Invariant Representation Research Articles

Optimal multi-kernel SVM classifier with rotation, illumination and scale invariant hybrid DWT-Shearlet based GLCM feature descriptor and its application to face recognition

In computer vision, we must handle with the various structural aspects of image or video data. The texture is one of the most important aspects of this type of data, which is utilised to identify objects or regions of interest in an image. As imaging conditions change, textures inside actual images significantly change in brightness, contrast, size, and skew. To recognise textures in real-world images, a similarity measure that is invariant to these features must be used. Existing recognition techniques did not perform well due to issues such as illumination, scale, and subject rotation. To address this issue, invariant feature representation methods are being developed to generate features that are insensitive to such variations. In this paper, we proposed a robust hybrid feature descriptor and predicted the faces under illumination, scale, and pose variations using an optimum multi-kernel support vector machine. Additionally, the suggested robust hybrid feature descriptor is enhanced by combining a hybrid transform composed of discrete wavelet and discrete shearlet transforms with some image statistical textural data. The proposed face recognition system is implemented in MATLAB, and analysed using various parameters to show proposed methods improved performance as compared to the state of the art methods.

Domain generalization in nematode classification

Nematode images captured by different microscopes may appear differently in terms of image color and image quality, resulting in these images laying in different learning domains. This can negatively impact nematode classification via deep learning. In this paper, we propose a local structure invariance guided (LSIG) domain generalization approach to enhance the model generalization of nematode local regions in unseen domains. First, a style transfer method is introduced to synthesize new domain image samples from the source domain. Unlike in the original input images, the color information of the synthetic images is changed, but their structural information is retained. Then, a metric learning strategy is designed to determine the cross-domain invariant structural representation between the source and new domains by pairwise learning. Each class is then effectively clustered, and a better decision boundary is determined to improve the model generalization. Overall, we demonstrate the effectiveness and robustness of the method on binary-class and multi-class classification tasks on diverse nematode datasets.

Unsupervised domain adaptation for the detection of cardiomegaly in cross-domain chest X-ray images.

In recent years, several deep learning approaches have been successfully applied in the field of medical image analysis. More specifically, different deep neural network architectures have been proposed and assessed for the detection of various pathologies based on chest X-ray images. While the performed assessments have shown very promising results, most of them consist in training and evaluating the performance of the proposed approaches on a single data set. However, the generalization of such models is quite limited in a cross-domain setting, since a significant performance degradation can be observed when these models are evaluated on data sets stemming from different medical centers or recorded under different protocols. The performance degradation is mostly caused by the domain shift between the training set and the evaluation set. To alleviate this problem, different unsupervised domain adaptation approaches are proposed and evaluated in the current work, for the detection of cardiomegaly based on chest X-ray images, in a cross-domain setting. The proposed approaches generate domain invariant feature representations by adapting the parameters of a model optimized on a large set of labeled samples, to a set of unlabeled images stemming from a different data set. The performed evaluation points to the effectiveness of the proposed approaches, since the adapted models outperform optimized models which are directly applied to the evaluation sets without any form of domain adaptation.

Optimizing training trajectories in variational autoencoders via latent Bayesian optimization approach *

Unsupervised and semi-supervised ML methods such as variational autoencoders (VAE) have become widely adopted across multiple areas of physics, chemistry, and materials sciences due to their capability in disentangling representations and ability to find latent manifolds for classification and/or regression of complex experimental data. Like other ML problems, VAEs require hyperparameter tuning, e.g. balancing the Kullback–Leibler and reconstruction terms. However, the training process and resulting manifold topology and connectivity depend not only on hyperparameters, but also their evolution during training. Because of the inefficiency of exhaustive search in a high-dimensional hyperparameter space for the expensive-to-train models, here we have explored a latent Bayesian optimization (zBO) approach for the hyperparameter trajectory optimization for the unsupervised and semi-supervised ML and demonstrated for joint-VAE with rotational invariances. We have demonstrated an application of this method for finding joint discrete and continuous rotationally invariant representations for modified national institute of standards and technology database (MNIST) and experimental data of a plasmonic nanoparticles material system. The performance of the proposed approach has been discussed extensively, where it allows for any high dimensional hyperparameter trajectory optimization of other ML models.

Variability in training unlocks generalization in visual perceptual learning through invariant representations

Stimulus and location specificity are long considered hallmarks of visual perceptual learning. This renders visual perceptual learning distinct from other forms of learning, where generalization can be more easily attained, and therefore unsuitable for practical applications, where generalization is key. Based on the hypotheses derived from the structure of the visual system, we test here whether stimulus variability can unlock generalization in perceptual learning. We train subjects in orientation discrimination, while we vary the amount of variability in a task-irrelevant feature, spatial frequency. We find that, independently of task difficulty, this manipulation enables generalization of learning to new stimuli and locations, while not negatively affecting the overall amount of learning on the task. We then use deep neural networks to investigate how variability unlocks generalization. We find that networks develop invariance to the task-irrelevant feature when trained with variable inputs. The degree of learned invariance strongly predicts generalization. A reliance on invariant representations can explain variability-induced generalization in visual perceptual learning. This suggests new targets for understanding the neural basis of perceptual learning in the higher-order visual cortex and presents an easy-to-implement modification of common training paradigms that may benefit practical applications.

Constructing Collective Variables Using Invariant Learned Representations.

On the time scales accessible to atomistic numerical modeling, chemical reactions are considered rare events. Therefore, the atomistic simulations are commonly biased along a low-dimensional representation of a chemical reaction in an atomic structure space, i.e., along the collective variables. However, suitable collective variables are often complicated to guess a priori. We propose a novel method of collective variable discovery based on dimensionality reduction of the atomic representation vectors. These linear-scaling and invariant representations can be either fixed (untrained) or learned by supervised training of the end-to-end machine learning potential. The learned representations are expected to reflect not only the structural but also the energetic features of the system that are transferable to all of the reactive transformation covered by the machine learning potential. We demonstrate our approach on four high-barrier reactions ranging from a simple gas-phase hydrogen jump reaction to complex reactions in periodic models of industrially relevant heterogeneous catalysts. High data efficiency, automatized feature extraction, favorable scaling, and retention of inherent invariances are all properties that are expected to enable fast and largely automatic construction of suitable collective variables even in highly complex reactive scenarios such as reactive/catalytic transformations at solid-liquid interfaces.

Improving the validity of neuroimaging decoding tests of invariant and configural neural representation.

Many research questions in sensory neuroscience involve determining whether the neural representation of a stimulus property is invariant or specific to a particular stimulus context (e.g., Is object representation invariant to translation? Is the representation of a face feature specific to the context of other face features?). Between these two extremes, representations may also be context-tolerant or context-sensitive. Most neuroimaging studies have used operational tests in which a target property is inferred from a significant test against the null hypothesis of the opposite property. For example, the popular cross-classification test concludes that representations are invariant or tolerant when the null hypothesis of specificity is rejected. A recently developed neurocomputational theory suggests two insights regarding such tests. First, tests against the null of context-specificity, and for the alternative of context-invariance, are prone to false positives due to the way in which the underlying neural representations are transformed into indirect measurements in neuroimaging studies. Second, jointly performing tests against the nulls of invariance and specificity allows one to reach more precise and valid conclusions about the underlying representations, particularly when the null of invariance is tested using the fine-grained information from classifier decision variables rather than only accuracies (i.e., using the decoding separability test). Here, we provide empirical and computational evidence supporting both of these theoretical insights. In our empirical study, we use encoding of orientation and spatial position in primary visual cortex as a case study, as previous research has established that these properties are encoded in a context-sensitive way. Using fMRI decoding, we show that the cross-classification test produces false-positive conclusions of invariance, but that more valid conclusions can be reached by jointly performing tests against the null of invariance. The results of two simulations further support both of these conclusions. We conclude that more valid inferences about invariance or specificity of neural representations can be reached by jointly testing against both hypotheses, and using neurocomputational theory to guide the interpretation of results.

Advancing 3D medical image analysis with variable dimension transform based supervised 3D pre-training

The difficulties in both data acquisition and annotation substantially restrict the sample sizes of training datasets for 3D medical imaging applications. Therefore, it is non-trivial to build well-performing 3D convolutional neural networks from scratch. Previous efforts on 3D pre-training have frequently relied on self-supervised approaches, which use either predictive or contrastive learning on unlabeled data to build invariant 3D representations. However, because of the unavailability of large-scale supervision information, obtaining semantically invariant and discriminative representations from these learning frameworks remains problematic. In this paper, we revisit an innovative yet simple fully-supervised 3D network pre-training framework to take advantage of semantic supervision from large-scale 2D natural image datasets. With a redesigned 3D network architecture, reformulated natural images are used to address the problem of data scarcity and develop powerful 3D representations. Comprehensive experiments on five benchmark datasets demonstrate that the proposed pre-trained models can effectively accelerate convergence while also improving accuracy for a variety of 3D medical imaging tasks such as classification, segmentation, and detection. In addition, as compared to training from scratch, it can save up to 60% of annotation efforts. On the NIH DeepLesion dataset, it also achieves state-of-the-art detection performance, outperforming earlier self-supervised and fully-supervised pre-training approaches, as well as methods that do training from scratch. To facilitate further development of 3D medical models, our code and pre-trained model weights are publicly available at https://github.com/urmagicsmine/CSPR.

Meta-Learning Based Tasks Similarity Representation for Cross Domain Lifelong Learning

Deep neural networks perform better in most specific single tasks than humans, but it is hard to handle a sequence of new tasks from different domains. The deep learning-based models always need to remember the parameters of the learned tasks to perform well in the new tasks and forfeit the ability to generalize from previous data, which is inconsistent with human learning. We propose a novel lifelong learning framework that can guide the model to learn new knowledge without forgetting the old knowledge through learning the similarity representation based on meta-learning. Specifically, we employ a cross-domain triplets network (CDTN) by minimizing the maximum mean discrepancy (MMD) between the current task and the knowledge base to learn the domain invariant similarity representation between tasks in different domains. Furthermore, we add a self-attention module to enhance the extraction of similarity features. Secondly, a spatial attention network (SAN) which can assign different weights according to the learned similarity representation of tasks, is proposed in addition. The experimental results show that our method effectively reduces catastrophic forgetting compared with the state-of-the-art methods when learning many tasks. Moreover, we show that the proposed method can hardly forget the old knowledge while continuously enhancing the performance of the old tasks, which is more in line with the human way of learning.

Adaptive Class Center Generalization Network: A Sparse Domain-Regressive Framework for Bearing Fault Diagnosis Under Unknown Working Conditions

Fault diagnosis is essential to ensure the bearing safety in the smart manufacturing. As the rotating bearings usually work under the variable working conditions, there may exist the differences between the data distributions of the training and test domains. Domain adaptation fault diagnosis (DAFD) has been adopted to handle with this domain shift phenomenon. But DAFD relies on the target domain heavily during its training process, while the target domain is always unavailable in real-world scenarios. To handle with this situation, this paper proposed a new adaptive class center generalization network (ACCGN). ACCGN is used to learn invariant feature representations of orientation signals from multiple source domains. First, ACCGN is used to learn the discriminative invariant fault feature from multi-source domains, and it combines the sparse domain regression framework and central loss to optimize the data features from inter- and intra-class simultaneously. Second, a new adaptive method is proposed to update the center in central loss and it can diminish the effect on the initialization center location. Third, a sparse domain regression framework is used to learn the inter-class invariant features. The proposed ACCGN has been tested on two famous bearing datasets, and the results have shown the effectiveness of the proposed ACCGN on the CWRU and JNU datasets.

Robust Low-Rank Clustering Contrastive Learning Integrating Transformer for Noisy Industrial Soft Sensors

Strong process noise and disturbance caused by operating conditions and internal systems have brought a considerable challenge for soft sensor modeling in industrial processes. Existing modeling methods mainly focus on removing the noise from the process data, which do not improve the accuracy of the soft sensor model effectively for it is unrealistic to get completely noise-free data in actual industrial processes. Therefore, a novel robust Low-rank Clustering Contrastive Learning (LrCCL) integrating Transformer (LrCCL-T) is proposed in this paper. Based on the data augmentation technology, the LrCCL is designed to learn intrinsic and invariant feature representations from the process data by combining low-rank prior (Lr) and adaptive clustering contrastive learning (CCL). The CCL can bring closer the pairs from the same sample and cluster. Moreover, the Lr which assumes that samples in the same cluster should lie in a low-dimensional subspace is utilized to enhance the learned feature representations by adding a low-rank constraint. Then, the Transformer is used to build the soft sensor model, which can extract dynamic temporal relationship between the learned feature representations and outputs. Finally, to verify the effectiveness and robustness of the proposed method, a public industrial thermal power dataset and an actual industrial polypropylene dataset are utilized to build the steam volume (SV) and the melt index (MI) soft sensor model, respectively. The contrastive and ablation experiment results show that the proposed LrCCL-T can achieve comparable accuracy to other state-of-the-art soft sensor methods under strong noise industrial environment.

Failure to Achieve Domain Invariance With Domain Generalization Algorithms: An Analysis in Medical Imaging

One prominent issue in the application of deep learning is the failure to generalize to data that lies on a different distribution to the training data. While many methods have been proposed to address this, prior work has shown that when operating under the same conditions most algorithms perform almost equally. As such, more work needs to be done to validate past and future methods before they are put into important scenarios like medical imaging. Our work analyses eight domain generalization algorithms across four important medical imaging classification datasets along with three standard natural image classification problems to discover the differences in how these methods operate in these different contexts. We assess these algorithms in terms of generalization capability, domain invariance, and representational sensitivity. Through this, we show that despite the differences between domain and content variations between natural and medical imaging there is little deviation in the operation of each method between natural images and medical images. Additionally, we show that all tested algorithms retain significant amounts of domain-specific information in their feature representations despite explicit training to remove it. Thus, revealing the failure point of all these methods is a lack of class-discriminative features extracted from out-of-distribution data. While these results show that methods that work well on natural imaging work similarly in medical imaging, no method outperforms baseline methods, highlighting the continuing gap of achieving adequate domain generalization. Similarly, the results also question the efficacy of optimizing for domain invariant representations as a method for generalizing to unseen domains.

From Laboratory to Field: Unsupervised Domain Adaptation for Plant Disease Recognition in the Wild.

Plant disease recognition is of vital importance to monitor plant development and predicting crop production. However, due to data degradation caused by different conditions of image acquisition, e.g., laboratory vs. field environment, machine learning-based recognition models generated within a specific dataset (source domain) tend to lose their validity when generalized to a novel dataset (target domain). To this end, domain adaptation methods can be leveraged for the recognition by learning invariant representations across domains. In this paper, we aim at addressing the issues of domain shift existing in plant disease recognition and propose a novel unsupervised domain adaptation method via uncertainty regularization, namely, Multi-Representation Subdomain Adaptation Network with Uncertainty Regularization for Cross-Species Plant Disease Classification (MSUN). Our simple but effective MSUN makes a breakthrough in plant disease recognition in the wild by using a large amount of unlabeled data and via nonadversarial training. Specifically, MSUN comprises multirepresentation, subdomain adaptation modules and auxiliary uncertainty regularization. The multirepresentation module enables MSUN to learn the overall structure of features and also focus on capturing more details by using the multiple representations of the source domain. This effectively alleviates the problem of large interdomain discrepancy. Subdomain adaptation is used to capture discriminative properties by addressing the issue of higher interclass similarity and lower intraclass variation. Finally, the auxiliary uncertainty regularization effectively suppresses the uncertainty problem due to domain transfer. MSUN was experimentally validated to achieve optimal results on the PlantDoc, Plant-Pathology, Corn-Leaf-Diseases, and Tomato-Leaf-Diseases datasets, with accuracies of 56.06%, 72.31%, 96.78%, and 50.58%, respectively, surpassing other state-of-the-art domain adaptation techniques considerably.

Learning Invariant Representations Under General Interventions on the Response

Learning Invariant Representations Under General Interventions on the Response

Learning Domain Invariant Representations for Generalizable Person Re-Identification.

Generalizable person Re-Identification (ReID) aims to learn ready-to-use cross-domain representations for direct cross-data evaluation, which has attracted growing attention in the recent computer vision (CV) community. In this work, we construct a structural causal model (SCM) among identity labels, identity-specific factors (clothing/shoes color etc.), and domain-specific factors (background, viewpoints etc.). According to the causal analysis, we propose a novel Domain Invariant Representation Learning for generalizable person Re-Identification (DIR-ReID) framework. Specifically, we propose to disentangle the identity-specific and domain-specific factors into two independent feature spaces, based on which an effective backdoor adjustment approximate implementation is proposed for serving as a causal intervention towards the SCM. Extensive experiments have been conducted, showing that DIR-ReID outperforms state-of-the-art (SOTA) methods on large-scale domain generalization (DG) ReID benchmarks.

Single-source Domain Expansion Network for Cross-Scene Hyperspectral Image Classification.

Currently, cross-scene hyperspectral image (HSI) classification has drawn increasing attention. It is necessary to train a model only on source domain (SD) and directly transferring the model to target domain (TD), when TD needs to be processed in real time and cannot be reused for training. Based on the idea of domain generalization, a Single-source Domain Expansion Network (SDEnet) is developed to ensure the reliability and effectiveness of domain extension. The method uses generative adversarial learning to train in SD and test in TD. A generator including semantic encoder and morph encoder is designed to generate the extended domain (ED) based on encoder-randomization-decoder architecture, where spatial randomization and spectral randomization are specifically used to generate variable spatial and spectral information, and the morphological knowledge is implicitly applied as domain invariant information during domain expansion. Furthermore, the supervised contrastive learning is employed in the discriminator to learn class-wise domain invariant representation, which drives intra-class samples of SD and ED. Meanwhile, adversarial training is designed to optimize the generator to drive intra-class samples of SD and ED to be separated. Extensive experiments on two public HSI datasets and one additional multispectral image (MSI) dataset demonstrate the superiority of the proposed method when compared with state-of-the-art techniques. The codes will be available from the website: https://github.com/YuxiangZhang-BIT/IEEE_TIP_SDEnet.

Language-Aware Domain Generalization Network for Cross-Scene Hyperspectral Image Classification

Text information including extensive prior knowledge about land cover classes has been ignored in hyperspectral image classification (HSI) tasks. It is necessary to explore the effectiveness of linguistic mode in assisting HSI classification. In addition, the large-scale pre-training image-text foundation models have demonstrated great performance in a variety of downstream applications, including zero-shot transfer. However, most domain generalization methods have never addressed mining linguistic modal knowledge to improve the generalization performance of model. To compensate for the inadequacies listed above, a Language-aware Domain Generalization Network (LDGnet) is proposed to learn cross-domain invariant representation from cross-domain shared prior knowledge. The proposed method only trains on the source domain (SD) and then transfers the model to the target domain (TD). The dual-stream architecture including image encoder and text encoder is used to extract visual and linguistic features, in which coarse-grained and fine-grained text representations are designed to extract two levels of linguistic features. Furthermore, linguistic features are used as cross-domain shared semantic space, and visual-linguistic alignment is completed by supervised contrastive learning in semantic space. Extensive experiments on three datasets demonstrate the superiority of the proposed method when compared with state-of-the-art techniques.

Cumulative foot pressure image recognition via Gabor filters and sparse representation classifier

Analysis of cumulative foot pressure images (CFPIs) can be utilised like other biometrics for personal recognition. This biometric, along with other biometrics such as gait, can be helpful in identification. Also, this overcomes some drawbacks of common biometrics (such as face and gait) in data recording and taking people's foot data without knowing them. In the process of capturing cumulative foot pressure images, the spatial and temporal changes of ground reaction force during one gait cycle are recorded. In this biometric, there are some challenges such as walking at different speeds. In this paper, we present a new approach based on Gabor filters as a feature vector and sparse representation classification (SRC). To reduce the feature dimension, eigenfoot and linear discriminative analysis (LDA) are also used. To obtain translation and rotation invariant representation, a normalisation preprocess is applied. We evaluate the proposed approach via dataset D of the Chinese Academy of Science (CASIA) Gait-Footprint dataset containing the cumulative foot pressure images of 88 persons. The experimental results indicate that the proposed method has higher accuracy than other tested methods, and also is robust to changes in walking speed.

Corrections to “Negation Invariant Representations of 3-D Vectors for Deep Learning Models Applied to Fault Geometry Mapping in 3-D Seismic Reflection Data”

Corrections to “Negation Invariant Representations of 3-D Vectors for Deep Learning Models Applied to Fault Geometry Mapping in 3-D Seismic Reflection Data”

Contrastive Learning for Radar HRRP Recognition With Missing Aspects

High resolution range profile (HRRP) has attracted increasing attention in radar automatic target recognition (RATR). However, the target-aspect missing problem in non-cooperative targets recognition, which is one of the most challenging tasks in RATR, has received very few contributions recently. The proposed work is motivated by a very simple observation, i.e., as compared with HRRP signals of interesting targets, sufficient unlabeled HRRP signals are much easier to acquire. However, these signals are often neglected since there is no label information. This work focuses on the target-aspect missing problem by using a dual self-supervised contrastive learning framework. The proposed model takes advantage of massive unlabeled HRRP signals to enhance the generalization ability. Specifically, we employ self-supervised contrastive learning and online clustering module for extracting target-aspect invariant representations. Additionally, a data augmentation strategy is used as a simple yet effective module. The experimental results demonstrate that the proposed method with fewer labels enhances the recognition performance as compared to supervised and self-supervised methods in missing aspect problem.