In the compressive spectral imaging (CSI) framework, different architectures have been proposed to recover high-resolution spectral images from compressive measurements. Since CSI architectures compactly capture the relevant information of the spectral image, various methods that extract classification features from compressive samples have been recently proposed. However, these techniques require a feature extraction procedure that reorders measurements using the information embedded in the coded aperture patterns. In this paper, a method that fuses features directly from multiresolution compressive measurements is proposed for spectral image classification. More precisely, the fusion method is formulated as an inverse problem that estimates high-spatial-resolution and low-dimensional feature bands from compressive measurements. To this end, the decimation matrices that describe the compressive measurements as degraded versions of the fused features are mathematically modeled using the information embedded in the coded aperture patterns. Furthermore, we include both a sparsity-promoting and a total-variation (TV) regularization terms to the fusion problem in order to consider the correlations between neighbor pixels, and therefore, improve the accuracy of pixel-based classifiers. To solve the fusion problem, we describe an algorithm based on the accelerated variant of the alternating direction method of multipliers (accelerated-ADMM). Additionally, a classification approach that includes the developed fusion method and a multilayer neural network is introduced. Finally, the proposed approach is evaluated on three remote sensing spectral images and a set of compressive measurements captured in the laboratory. Extensive simulations show that the proposed classification approach outperforms other approaches under various performance metrics.
Few-shot learning for fine-grained image classification has gained recent attention in computer vision. Among the approaches for few-shot learning, due to the simplicity and effectiveness, metric-based methods are favorably state-of-the-art on many tasks. Most of the metric-based methods assume a single similarity measure and thus obtain a single feature space. However, if samples can simultaneously be well classified via two distinct similarity measures, the samples within a class can distribute more compactly in a smaller feature space, producing more discriminative feature maps. Motivated by this, we propose a so-called \textit{Bi-Similarity Network} (\textit{BSNet}) that consists of a single embedding module and a bi-similarity module of two similarity measures. After the support images and the query images pass through the convolution-based embedding module, the bi-similarity module learns feature maps according to two similarity measures of diverse characteristics. In this way, the model is enabled to learn more discriminative and less similarity-biased features from few shots of fine-grained images, such that the model generalization ability can be significantly improved. Through extensive experiments by slightly modifying established metric/similarity based networks, we show that the proposed approach produces a substantial improvement on several fine-grained image benchmark datasets. Codes are available at: https://github.com/spraise/BSNet
Meta-learning has become a practical approach towards few-shot image classification, where "a strategy to learn a classifier" is meta-learned on labeled base classes and can be applied to tasks with novel classes. We remove the requirement of base class labels and learn generalizable embeddings via Unsupervised Meta-Learning (UML). Specifically, episodes of tasks are constructed with data augmentations from unlabeled base classes during meta-training, and we apply embedding-based classifiers to novel tasks with labeled few-shot examples during meta-test. We observe two elements play important roles in UML, i.e., the way to sample tasks and measure similarities between instances. Thus we obtain a strong baseline with two simple modifications -- a sufficient sampling strategy constructing multiple tasks per episode efficiently together with a semi-normalized similarity. We then take advantage of the characteristics of tasks from two directions to get further improvements. First, synthesized confusing instances are incorporated to help extract more discriminative embeddings. Second, we utilize an additional task-specific embedding transformation as an auxiliary component during meta-training to promote the generalization ability of the pre-adapted embeddings. Experiments on few-shot learning benchmarks verify that our approaches outperform previous UML methods and achieve comparable or even better performance than its supervised variants.
The width of a neural network matters since increasing the width will necessarily increase the model capacity. However, the performance of a network does not improve linearly with the width and soon gets saturated. In this case, we argue that increasing the number of networks (ensemble) can achieve better accuracy-efficiency trade-offs than purely increasing the width. To prove it, one large network is divided into several small ones regarding its parameters and regularization components. Each of these small networks has a fraction of the original one's parameters. We then train these small networks together and make them see various views of the same data to increase their diversity. During this co-training process, networks can also learn from each other. As a result, small networks can achieve better ensemble performance than the large one with few or no extra parameters or FLOPs, \ie, achieving better accuracy-efficiency trade-offs. Small networks can also achieve faster inference speed than the large one by concurrent running. All of the above shows that the number of networks is a new dimension of model scaling. We validate our argument with 8 different neural architectures on common benchmarks through extensive experiments. The code is available at \url{https://github.com/FreeformRobotics/Divide-and-Co-training}.
Self-supervised visual representation learning has seen huge progress recently, but no large scale evaluation has compared the many models now available. We evaluate the transfer performance of 13 top self-supervised models on 40 downstream tasks, including many-shot and few-shot recognition, object detection, and dense prediction. We compare their performance to a supervised baseline and show that on most tasks the best self-supervised models outperform supervision, confirming the recently observed trend in the literature. We find ImageNet Top-1 accuracy to be highly correlated with transfer to many-shot recognition, but increasingly less so for few-shot, object detection and dense prediction. No single self-supervised method dominates overall, suggesting that universal pre-training is still unsolved. Our analysis of features suggests that top self-supervised learners fail to preserve colour information as well as supervised alternatives, but tend to induce better classifier calibration, and less attentive overfitting than supervised learners.
Virtual Adversarial Training (VAT) has shown impressive results among recently developed regularization methods called consistency regularization. VAT utilizes adversarial samples, generated by injecting perturbation in the input space, for training and thereby enhances the generalization ability of a classifier. However, such adversarial samples can be generated only within a very small area around the input data point, which limits the adversarial effectiveness of such samples. To address this problem we propose LVAT (Latent space VAT), which injects perturbation in the latent space instead of the input space. LVAT can generate adversarial samples flexibly, resulting in more adverse effects and thus more effective regularization. The latent space is built by a generative model, and in this paper, we examine two different type of models: variational auto-encoder and normalizing flow, specifically Glow. We evaluated the performance of our method in both supervised and semi-supervised learning scenarios for an image classification task using SVHN and CIFAR-10 datasets. In our evaluation, we found that our method outperforms VAT and other state-of-the-art methods.
Few-shot learning (FSL) approaches are usually based on an assumption that the pre-trained knowledge can be obtained from base (seen) categories and can be well transferred to novel (unseen) categories. However, there is no guarantee, especially for the latter part. This issue leads to the unknown nature of the inference process in most FSL methods, which hampers its application in some risk-sensitive areas. In this paper, we reveal a new way to perform FSL for image classification, using visual representations from the backbone model and weights generated by a newly-emerged explainable classifier. The weighted representations only include a minimum number of distinguishable features and the visualized weights can serve as an informative hint for the FSL process. Finally, a discriminator will compare the representations of each pair of the images in the support set and the query set. Pairs with the highest scores will decide the classification results. Experimental results prove that the proposed method can achieve both good accuracy and satisfactory explainability on three mainstream datasets.
In real-world classification tasks, each class often comprises multiple finer-grained "subclasses." As the subclass labels are frequently unavailable, models trained using only the coarser-grained class labels often exhibit highly variable performance across different subclasses. This phenomenon, known as hidden stratification, has important consequences for models deployed in safety-critical applications such as medicine. We propose GEORGE, a method to both measure and mitigate hidden stratification even when subclass labels are unknown. We first observe that unlabeled subclasses are often separable in the feature space of deep neural networks, and exploit this fact to estimate subclass labels for the training data via clustering techniques. We then use these approximate subclass labels as a form of noisy supervision in a distributionally robust optimization objective. We theoretically characterize the performance of GEORGE in terms of the worst-case generalization error across any subclass. We empirically validate GEORGE on a mix of real-world and benchmark image classification datasets, and show that our approach boosts worst-case subclass accuracy by up to 22 percentage points compared to standard training techniques, without requiring any prior information about the subclasses.
While knowledge distillation (transfer) has been attracting attentions from the research community, the recent development in the fields has heightened the need for reproducible studies and highly generalized frameworks to lower barriers to such high-quality, reproducible deep learning research. Several researchers voluntarily published frameworks used in their knowledge distillation studies to help other interested researchers reproduce their original work. Such frameworks, however, are usually neither well generalized nor maintained, thus researchers are still required to write a lot of code to refactor/build on the frameworks for introducing new methods, models, datasets and designing experiments. In this paper, we present our developed open-source framework built on PyTorch and dedicated for knowledge distillation studies. The framework is designed to enable users to design experiments by declarative PyYAML configuration files, and helps researchers complete the recently proposed ML Code Completeness Checklist. Using the developed framework, we demonstrate its various efficient training strategies, and implement a variety of knowledge distillation methods. We also reproduce some of their original experimental results on the ImageNet and COCO datasets presented at major machine learning conferences such as ICLR, NeurIPS, CVPR and ECCV, including recent state-of-the-art methods. All the source code, configurations, log files and trained model weights are publicly available at https://github.com/yoshitomo-matsubara/torchdistill .
In open-world learning, an agent starts with a set of known classes, detects, and manages things that it does not know, and learns them over time from a non-stationary stream of data. Open-world learning is related to but also distinct from a multitude of other learning problems and this paper briefly analyzes the key differences between a wide range of problems including incremental learning, generalized novelty discovery, and generalized zero-shot learning. This paper formalizes various open-world learning problems including open-world learning without labels. These open-world problems can be addressed with modifications to known elements, we present a new framework that enables agents to combine various modules for novelty-detection, novelty-characterization, incremental learning, and instance management to learn new classes from a stream of unlabeled data in an unsupervised manner, survey how to adapt a few state-of-the-art techniques to fit the framework and use them to define seven baselines for performance on the open-world learning without labels problem. We then discuss open-world learning quality and analyze how that can improve instance management. We also discuss some of the general ambiguity issues that occur in open-world learning without labels.
Softmax function is widely used in artificial neural networks for multiclass classification, multilabel classification, attention mechanisms, etc. However, its efficacy is often questioned in literature. The log-softmax loss has been shown to belong to a more generic class of loss functions, called spherical family, and its member log-Taylor softmax loss is arguably the best alternative in this class. In another approach which tries to enhance the discriminative nature of the softmax function, soft-margin softmax (SM-softmax) has been proposed to be the most suitable alternative. In this work, we investigate Taylor softmax, SM-softmax and our proposed SM-Taylor softmax, an amalgamation of the earlier two functions, as alternatives to softmax function. Furthermore, we explore the effect of expanding Taylor softmax up to ten terms (original work proposed expanding only to two terms) along with the ramifications of considering Taylor softmax to be a finite or infinite series during backpropagation. Our experiments for the image classification task on different datasets reveal that there is always a configuration of the SM-Taylor softmax function that outperforms the normal softmax function and its other alternatives.
Today's deep models are often unable to detect inputs which do not belong to the training distribution. This gives rise to confident incorrect predictions which could lead to devastating consequences in many important application fields such as healthcare and autonomous driving. Interestingly, both discriminative and generative models appear to be equally affected. Consequently, this vulnerability represents an important research challenge. We consider an outlier detection approach based on discriminative training with jointly learned synthetic outliers. We obtain the synthetic outliers by sampling an RNVP model which is jointly trained to generate datapoints at the border of the training distribution. We show that this approach can be adapted for simultaneous semantic segmentation and dense outlier detection. We present image classification experiments on CIFAR-10, as well as semantic segmentation experiments on three existing datasets (StreetHazards, WD-Pascal, Fishyscapes Lost & Found), and one contributed dataset. Our models perform competitively with respect to the state of the art despite producing predictions with only one forward pass.
To collect large scale annotated data, it is inevitable to introduce label noise, i.e., incorrect class labels. To be robust against label noise, many successful methods rely on the noisy classifiers (i.e., models trained on the noisy training data) to determine whether a label is trustworthy. However, it remains unknown why this heuristic works well in practice. In this paper, we provide the first theoretical explanation for these methods. We prove that the prediction of a noisy classifier can indeed be a good indicator of whether the label of a training data is clean. Based on the theoretical result, we propose a novel algorithm that corrects the labels based on the noisy classifier prediction. The corrected labels are consistent with the true Bayesian optimal classifier with high probability. We incorporate our label correction algorithm into the training of deep neural networks and train models that achieve superior testing performance on multiple public datasets.
Contrastive learning methods have significantly narrowed the gap between supervised and unsupervised learning on computer vision tasks. In this paper, we explore their application to geo-located datasets, e.g. remote sensing, where unlabeled data is often abundant but labeled data is scarce. We first show that due to their different characteristics, a non-trivial gap persists between contrastive and supervised learning on standard benchmarks. To close the gap, we propose novel training methods that exploit the spatio-temporal structure of remote sensing data. We leverage spatially aligned images over time to construct temporal positive pairs in contrastive learning and geo-location to design pre-text tasks. Our experiments show that our proposed method closes the gap between contrastive and supervised learning on image classification, object detection and semantic segmentation for remote sensing. Moreover, we demonstrate that the proposed method can also be applied to geo-tagged ImageNet images, improving downstream performance on various tasks. Project Webpage can be found at this link geography-aware-ssl.github.io.
We present a multigrid-in-channels (MGIC) approach that tackles the quadratic growth of the number of parameters with respect to the number of channels in standard convolutional neural networks (CNNs). Thereby our approach addresses the redundancy in CNNs that is also exposed by the recent success of lightweight CNNs. Lightweight CNNs can achieve comparable accuracy to standard CNNs with fewer parameters; however, the number of weights still scales quadratically with the CNN's width. Our MGIC architectures replace each CNN block with an MGIC counterpart that utilizes a hierarchy of nested grouped convolutions of small group size to address this. Hence, our proposed architectures scale linearly with respect to the network's width while retaining full coupling of the channels as in standard CNNs. Our extensive experiments on image classification, segmentation, and point cloud classification show that applying this strategy to different architectures like ResNet and MobileNetV3 reduces the number of parameters while obtaining similar or better accuracy.
We address the challenging problem of whole slide image (WSI) classification. WSIs have very high resolutions and usually lack localized annotations. WSI classification can be cast as a multiple instance learning (MIL) problem when only slide-level labels are available. We propose a MIL-based method for WSI classification and tumor detection that does not require localized annotations. Our method has three major components. First, we introduce a novel MIL aggregator that models the relations of the instances in a dual-stream architecture with trainable distance measurement. Second, since WSIs can produce large or unbalanced bags that hinder the training of MIL models, we propose to use self-supervised contrastive learning to extract good representations for MIL and alleviate the issue of prohibitive memory cost for large bags. Third, we adopt a pyramidal fusion mechanism for multiscale WSI features, and further improve the accuracy of classification and localization. Our model is evaluated on two representative WSI datasets. The classification accuracy of our model compares favorably to fully-supervised methods, with less than 2% accuracy gap across datasets. Our results also outperform all previous MIL-based methods. Additional benchmark results on standard MIL datasets further demonstrate the superior performance of our MIL aggregator on general MIL problems. GitHub repository: https://github.com/binli123/dsmil-wsi
Explanation methods facilitate the development of models that learn meaningful concepts and avoid exploiting spurious correlations. We illustrate a previously unrecognized limitation of the popular neural network explanation method Grad-CAM: as a side effect of the gradient averaging step, Grad-CAM sometimes highlights locations the model did not actually use. To solve this problem, we propose HiResCAM, a novel class-specific explanation method that is guaranteed to highlight only the locations the model used to make each prediction. We prove that HiResCAM is a generalization of CAM and explore the relationships between HiResCAM and other gradient-based explanation methods. Experiments on PASCAL VOC 2012, including crowd-sourced evaluations, illustrate that while HiResCAM's explanations faithfully reflect the model, Grad-CAM often expands the attention to create bigger and smoother visualizations. Overall, this work advances convolutional neural network explanation approaches and may aid in the development of trustworthy models for sensitive applications.
In this work, we propose CBiGAN -- a novel method for anomaly detection in images, where a consistency constraint is introduced as a regularization term in both the encoder and decoder of a BiGAN. Our model exhibits fairly good modeling power and reconstruction consistency capability. We evaluate the proposed method on MVTec AD -- a real-world benchmark for unsupervised anomaly detection on high-resolution images -- and compare against standard baselines and state-of-the-art approaches. Experiments show that the proposed method improves the performance of BiGAN formulations by a large margin and performs comparably to expensive state-of-the-art iterative methods while reducing the computational cost. We also observe that our model is particularly effective in texture-type anomaly detection, as it sets a new state of the art in this category. Our code is available at https://github.com/fabiocarrara/cbigan-ad/.
The recently introduced locally orderless tensor network (LoTeNet) for supervised image classification uses matrix product state (MPS) operations on grids of transformed image patches. The resulting patch representations are combined back together into the image space and aggregated hierarchically using multiple MPS blocks per layer to obtain the final decision rules. In this work, we propose a non-patch based modification to LoTeNet that performs one MPS operation per layer, instead of several patch-level operations. The spatial information in the input images to MPS blocks at each layer is squeezed into the feature dimension, similar to LoTeNet, to maximise retained spatial correlation between pixels when images are flattened into 1D vectors. The proposed multi-layered tensor network (MLTN) is capable of learning linear decision boundaries in high dimensional spaces in a multi-layered setting, which results in a reduction in the computation cost compared to LoTeNet without any degradation in performance.
Attention maps are a popular way of explaining the decisions of convolutional networks for image classification. Typically, for each image of interest, a single attention map is produced, which assigns weights to pixels based on their importance to the classification. A single attention map, however, provides an incomplete understanding since there are often many other maps that explain a classification equally well. In this paper, we introduce structured attention graphs (SAGs), which compactly represent sets of attention maps for an image by capturing how different combinations of image regions impact a classifier's confidence. We propose an approach to compute SAGs and a visualization for SAGs so that deeper insight can be gained into a classifier's decisions. We conduct a user study comparing the use of SAGs to traditional attention maps for answering counterfactual questions about image classifications. Our results show that the users are more correct when answering comparative counterfactual questions based on SAGs compared to the baselines.