Online Normalization is a new technique for normalizing the hidden activations of a neural network. Like Batch Normalization, it normalizes the sample dimension. While Online Normalization does not use batches, it is as accurate as Batch Normalization. We resolve a theoretical limitation of Batch Normalization by introducing an unbiased technique for computing the gradient of normalized activations. Online Normalization works with automatic differentiation by adding statistical normalization as a primitive. This technique can be used in cases not covered by some other normalizers, such as recurrent networks, fully connected networks, and networks with activation memory requirements prohibitive for batching. We show its applications to image classification, image segmentation, and language modeling. We present formal proofs and experimental results on ImageNet, CIFAR, and PTB datasets.
In most computer vision and image analysis problems, it is necessary to define a similarity measure between two or more different objects or images. Template matching is a classic and fundamental method used to score similarities between objects using certain mathematical algorithms. In this paper, we reviewed the basic concept of matching, as well as advances in template matching and applications such as invariant features or novel applications in medical image analysis. Additionally, deformable models and templates originating from classic template matching were discussed. These models have broad applications in image registration, and they are a fundamental aspect of novel machine vision or deep learning algorithms, such as convolutional neural networks (CNN), which perform shift and scale invariant functions followed by classification. In general, although template matching methods have restrictions which limit their application, they are recommended for use with other object recognition methods as pre- or post-processing steps. Combining a template matching technique such as normalized cross-correlation or dice coefficient with a robust decision-making algorithm yields a significant improvement in the accuracy rate for object detection and recognition.
Traditional survival models such as the Cox proportional hazards model are typically based on scalar or categorical clinical features. With the advent of increasingly large image datasets, it has become feasible to incorporate quantitative image features into survival prediction. So far, this kind of analysis is mostly based on radiomics features, i.e. a fixed set of features that is mathematically defined a priori. To capture highly abstract information, it is desirable to learn the feature extraction using convolutional neural networks. However, for tomographic medical images, model training is difficult because on the one hand, only few samples of 3D image data fit into one batch at once and on the other hand, survival loss functions are essentially ordering measures that require large batch sizes. In this work, we show that by simplifying survival analysis to median survival classification, convolutional neural networks can be trained with small batch sizes and learn features that predict survival equally well as end-to-end hazard prediction networks. Our approach outperforms the previous state of the art in a publicly available lung cancer dataset.
Self-supervised learning has been widely used to obtain transferrable representations from unlabeled images. Especially, recent contrastive learning methods have shown impressive performances on downstream image classification tasks. While these contrastive methods mainly focus on generating invariant global representations at the image-level under semantic-preserving transformations, they are prone to overlook spatial consistency of local representations and therefore have a limitation in pretraining for localization tasks such as object detection and instance segmentation. Moreover, aggressively cropped views used in existing contrastive methods can minimize representation distances between the semantically different regions of a single image. In this paper, we propose a spatially consistent representation learning algorithm (SCRL) for multi-object and location-specific tasks. In particular, we devise a novel self-supervised objective that tries to produce coherent spatial representations of a randomly cropped local region according to geometric translations and zooming operations. On various downstream localization tasks with benchmark datasets, the proposed SCRL shows significant performance improvements over the image-level supervised pretraining as well as the state-of-the-art self-supervised learning methods. Code is available at https://github.com/kakaobrain/scrl
Unsupervised learning has grown in popularity because of the difficulty of collecting annotated data and the development of modern frameworks that allow us to learn from unlabeled data. Existing studies, however, either disregard variations at different levels of similarity or only consider negative samples from one batch. We argue that image pairs should have varying degrees of similarity, and the negative samples should be allowed to be drawn from the entire dataset. In this work, we propose Search-based Unsupervised Visual Representation Learning (SUVR) to learn better image representations in an unsupervised manner. We first construct a graph from the image dataset by the similarity between images, and adopt the concept of graph traversal to explore positive samples. In the meantime, we make sure that negative samples can be drawn from the full dataset. Quantitative experiments on five benchmark image classification datasets demonstrate that SUVR can significantly outperform strong competing methods on unsupervised embedding learning. Qualitative experiments also show that SUVR can produce better representations in which similar images are clustered closer together than unrelated images in the latent space.
We propose a novel framework to address the real-world challenging task of Single Image Test Time Adaptation in an open and dynamic environment. We leverage large scale Vision Language Models like CLIP to enable real time adaptation on a per-image basis without access to source data or ground truth labels. Since the deployed model can also encounter unseen classes in an open world, we first employ a simple and effective Out of Distribution (OOD) detection module to distinguish between weak and strong OOD samples. We propose a novel contrastive learning based objective to enhance the discriminability between weak and strong OOD samples by utilizing small, dynamically updated feature banks. Finally, we also employ a classification objective for adapting the model using the reliable weak OOD samples. The proposed framework ROSITA combines these components, enabling continuous online adaptation of Vision Language Models on a single image basis. Extensive experimentation on diverse domain adaptation benchmarks validates the effectiveness of the proposed framework. Our code can be found at the project site https://manogna-s.github.io/rosita/
Identifying the authenticity and processing history of an image is an important task in multimedia forensics. By analyzing traces left by different image manipulations, researchers have been able to develop several algorithms capable of detecting targeted editing operations. While this approach has led to the development of several successful forensic algorithms, an important problem remains: creating forensic detectors for different image manipulations is a difficult and time consuming process. Furthermore, forensic analysts need ‘general purpose’ forensic algorithms capable of detecting multiple different image manipulations. In this paper, we address both of these problems by proposing a new general purpose forensic approach using convolutional neural networks (CNNs). While CNNs are capable of learning classification features directly from data, in their existing form they tend to learn features representative of an image’s content. To overcome this issue, we have developed a new type of CNN layer, called a constrained convolutional layer, that is able to jointly suppress an image’s content and adaptively learn manipulation detection features. Through a series of experiments, we show that our proposed constrained CNN is able to learn manipulation detection features directly from data. Our experimental results demonstrate that our CNN can detect multiple different editing operations with up to 99.97% accuracy and outperform the existing state-of-the-art general purpose manipulation detector. Furthermore, our constrained CNN can still accurately detect image manipulations in realistic scenarios where there is a source camera model mismatch between the training and testing data
We propose a novel multi-task learning architecture, which allows learning of task-specific feature-level attention. Our design, the Multi-Task Attention Network (MTAN), consists of a single shared network containing a global feature pool, together with a soft-attention module for each task. These modules allow for learning of task-specific features from the global features, whilst simultaneously allowing for features to be shared across different tasks. The architecture can be trained end-to-end and can be built upon any feed-forward neural network, is simple to implement, and is parameter efficient. We evaluate our approach on a variety of datasets, across both image-to-image predictions and image classification tasks. We show that our architecture is state-of-the-art in multi-task learning compared to existing methods, and is also less sensitive to various weighting schemes in the multi-task loss function. Code is available at https://github.com/lorenmt/mtan.
Human-Object Interaction (HOI) detection is a task of identifying "a set of interactions" in an image, which involves the i) localization of the subject (i.e., humans) and target (i.e., objects) of interaction, and ii) the classification of the interaction labels. Most existing methods have indirectly addressed this task by detecting human and object instances and individually inferring every pair of the detected instances. In this paper, we present a novel framework, referred to by HOTR, which directly predicts a set of <human, object, interaction> triplets from an image based on a transformer encoder-decoder architecture. Through the set prediction, our method effectively exploits the inherent semantic relationships in an image and does not require time-consuming post-processing which is the main bottleneck of existing methods. Our proposed algorithm achieves the state-of-the-art performance in two HOI detection benchmarks with an inference time under 1 ms after object detection.
In this paper, we propose QACE, a new metric based on Question Answering for Caption Evaluation. QACE generates questions on the evaluated caption and checks its content by asking the questions on either the reference caption or the source image. We first develop QACE-Ref that compares the answers of the evaluated caption to its reference, and report competitive results with the state-of-the-art metrics. To go further, we propose QACE-Img, which asks the questions directly on the image, instead of reference. A Visual-QA system is necessary for QACE-Img. Unfortunately, the standard VQA models are framed as a classification among only a few thousand categories. Instead, we propose Visual-T5, an abstractive VQA system. The resulting metric, QACE-Img is multi-modal, reference-less, and explainable. Our experiments show that QACE-Img compares favorably w.r.t. other reference-less metrics. We will release the pre-trained models to compute QACE.
Modern healthcare often utilises radiographic images alongside textual reports for diagnostics, encouraging the use of Vision-Language Self-Supervised Learning (VL-SSL) with large pre-trained models to learn versatile medical vision representations. However, most existing VL-SSL frameworks are trained end-to-end, which is computation-heavy and can lose vital prior information embedded in pre-trained encoders. To address both issues, we introduce the backbone-agnostic Adaptor framework, which preserves medical knowledge in pre-trained image and text encoders by keeping them frozen, and employs a lightweight Adaptor module for cross-modal learning. Experiments on medical image classification and segmentation tasks across three datasets reveal that our framework delivers competitive performance while cutting trainable parameters by over 90% compared to current pre-training approaches. Notably, when fine-tuned with just 1% of data, Adaptor outperforms several Transformer-based methods trained on full datasets in medical image segmentation.
Deep learning based approaches to Computer Aided Diagnosis (CAD) typically pose the problem as an image classification (Normal or Abnormal) problem. These systems achieve high to very high accuracy in specific disease detection for which they are trained but lack in terms of an explanation for the provided decision/classification result. The activation maps which correspond to decisions do not correlate well with regions of interest for specific diseases. This paper examines this problem and proposes an approach which mimics the clinical practice of looking for an evidence prior to diagnosis. A CAD model is learnt using a mixed set of information: class labels for the entire training set of images plus a rough localisation of suspect regions as an extra input for a smaller subset of training images for guiding the learning. The proposed approach is illustrated with detection of diabetic macular edema (DME) from OCT slices. Results of testing on on a large public dataset show that with just a third of images with roughly segmented fluid filled regions, the classification accuracy is on par with state of the art methods while providing a good explanation in the form of anatomically accurate heatmap /region of interest. The proposed solution is then adapted to Breast Cancer detection from mammographic images. Good evaluation results on public datasets underscores the generalisability of the proposed solution.
Classification of skull fracture is a challenging task for both radiologists and researchers. Skull fractures result in broken pieces of bone, which can cut into the brain and cause bleeding and other injury types. So it is vital to detect and classify the fracture very early. In real world, often fractures occur at multiple sites. This makes it harder to detect the fracture type where many fracture types might summarize a skull fracture. Unfortunately, manual detection of skull fracture and the classification process is time-consuming, threatening a patient's life. Because of the emergence of deep learning, this process could be automated. Convolutional Neural Networks (CNNs) are the most widely used deep learning models for image categorization because they deliver high accuracy and outstanding outcomes compared to other models. We propose a new model called SkullNetV1 comprising a novel CNN by taking advantage of CNN for feature extraction and lazy learning approach which acts as a classifier for classification of skull fractures from brain CT images to classify five fracture types. Our suggested model achieved a subset accuracy of 88%, an F1 score of 93%, the Area Under the Curve (AUC) of 0.89 to 0.98, a Hamming score of 92% and a Hamming loss of 0.04 for this seven-class multi-labeled classification.
Image understanding heavily relies on accurate multi-label classification. In recent years, deep learning algorithms have become very successful for such tasks, and various commercial and open-source APIs have been released for public use. However, these APIs are often trained on different datasets, which, besides affecting their performance, might pose a challenge to their performance evaluation. This challenge concerns the different object-class dictionaries of the APIs' training dataset and the benchmark dataset, in which the predicted labels are semantically similar to the benchmark labels but considered different simply because they have different wording in the dictionaries. To face this challenge, we propose semantic similarity metrics to obtain richer understating of the APIs predicted labels and thus their performance. In this study, we evaluate and compare the performance of 13 of the most prominent commercial and open-source APIs in a best-of-breed challenge on the Visual Genome and Open Images benchmark datasets. Our findings demonstrate that, while using traditional metrics, the Microsoft Computer Vision, Imagga, and IBM APIs performed better than others. However, applying semantic metrics also unveil the InceptionResNet-v2, Inception-v3, and ResNet50 APIs, which are trained only with the simple ImageNet dataset, as challengers for top semantic performers.
We consider the problem of representation learning for graph data. Convolutional neural networks can naturally operate on images, but have significant challenges in dealing with graph data. Given images are special cases of graphs with nodes lie on 2D lattices, graph embedding tasks have a natural correspondence with image pixel-wise prediction tasks such as segmentation. While encoder-decoder architectures like U-Nets have been successfully applied on many image pixel-wise prediction tasks, similar methods are lacking for graph data. This is due to the fact that pooling and up-sampling operations are not natural on graph data. To address these challenges, we propose novel graph pooling (gPool) and unpooling (gUnpool) operations in this work. The gPool layer adaptively selects some nodes to form a smaller graph based on their scalar projection values on a trainable projection vector. We further propose the gUnpool layer as the inverse operation of the gPool layer. The gUnpool layer restores the graph into its original structure using the position information of nodes selected in the corresponding gPool layer. Based on our proposed gPool and gUnpool layers, we develop an encoder-decoder model on graph, known as the graph U-Nets. Our experimental results on node classification and graph classification tasks demonstrate that our methods achieve consistently better performance than previous models.
Deep neural networks used for image classification often use convolutional filters to extract distinguishing features before passing them to a linear classifier. Most interpretability literature focuses on providing semantic meaning to convolutional filters to explain a model's reasoning process and confirm its use of relevant information from the input domain. Fully connected layers can be studied by decomposing their weight matrices using a singular value decomposition, in effect studying the correlations between the rows in each matrix to discover the dynamics of the map. In this work we define a singular value decomposition for the weight tensor of a convolutional layer, which provides an analogous understanding of the correlations between filters, exposing the dynamics of the convolutional map. We validate our definition using recent results in random matrix theory. By applying the decomposition across the linear layers of an image classification network we suggest a framework against which interpretability methods might be applied using hypergraphs to model class separation. Rather than looking to the activations to explain the network, we use the singular vectors with the greatest corresponding singular values for each linear layer to identify those features most important to the network. We illustrate our approach with examples and introduce the DeepDataProfiler library, the analysis tool used for this study.
In search of robust and generalizable machine learning models, Domain Generalization (DG) has gained significant traction during the past few years. The goal in DG is to produce models which continue to perform well when presented with data distributions different from the ones available during training. While deep convolutional neural networks (CNN) have been able to achieve outstanding performance on downstream computer vision tasks, they still often fail to generalize on previously unseen data Domains. Therefore, in this work we focus on producing a model which is able to remain robust under data distribution shift and propose an alternative regularization technique for convolutional neural network architectures in the single-source DG image classification setting. To mitigate the problem caused by domain shift between source and target data, we propose augmenting intermediate feature maps of CNNs. Specifically, we pass them through a novel Augmentation Layer} to prevent models from overfitting on the training set and improve their cross-domain generalization. To the best of our knowledge, this is the first paper proposing such a setup for the DG image classification setting. Experiments on the DG benchmark datasets of PACS, VLCS, Office-Home and TerraIncognita validate the effectiveness of our method, in which our model surpasses state-of-the-art algorithms in most cases.
The development and adoption of Vision Transformers and other deep-learning architectures for image classification tasks has been rapid. However, the "black box" nature of neural networks is a barrier to adoption in applications where explainability is essential. While some techniques for generating explanations have been proposed, primarily for Convolutional Neural Networks, adapting such techniques to the new paradigm of Vision Transformers is non-trivial. This paper presents T-TAME, Transformer-compatible Trainable Attention Mechanism for Explanations, a general methodology for explaining deep neural networks used in image classification tasks. The proposed architecture and training technique can be easily applied to any convolutional or Vision Transformer-like neural network, using a streamlined training approach. After training, explanation maps can be computed in a single forward pass; these explanation maps are comparable to or outperform the outputs of computationally expensive perturbation-based explainability techniques, achieving SOTA performance. We apply T-TAME to three popular deep learning classifier architectures, VGG-16, ResNet-50, and ViT-B-16, trained on the ImageNet dataset, and we demonstrate improvements over existing state-of-the-art explainability methods. A detailed analysis of the results and an ablation study provide insights into how the T-TAME design choices affect the quality of the generated explanation maps.
In medical imaging, a general problem is that it is costly and time consuming to collect high quality data from healthy and diseased subjects. Generative adversarial networks (GANs) is a deep learning method that has been developed for synthesizing data. GANs can thereby be used to generate more realistic training data, to improve classification performance of machine learning algorithms. Another application of GANs is image-to-image translations, e.g. generating magnetic resonance (MR) images from computed tomography (CT) images, which can be used to obtain multimodal datasets from a single modality. Here, we evaluate two unsupervised GAN models (CycleGAN and UNIT) for image-to-image translation of T1- and T2-weighted MR images, by comparing generated synthetic MR images to ground truth images. We also evaluate two supervised models; a modification of CycleGAN and a pure generator model. A small perceptual study was also performed to evaluate how visually realistic the synthesized images are. It is shown that the implemented GAN models can synthesize visually realistic MR images (incorrectly labeled as real by a human). It is also shown that models producing more visually realistic synthetic images not necessarily have better quantitative error measurements, when compared to ground truth data. Code is available at https://github.com/simontomaskarlsson/GAN-MRI
Historically, machine learning methods have not been designed with security in mind. In turn, this has given rise to adversarial examples, carefully perturbed input samples aimed to mislead detection at test time, which have been applied to attack spam and malware classification, and more recently to attack image classification. Consequently, an abundance of research has been devoted to designing machine learning methods that are robust to adversarial examples. Unfortunately, there are desiderata besides robustness that a secure and safe machine learning model must satisfy, such as fairness and privacy. Recent work by Song et al. (2019) has shown, empirically, that there exists a trade-off between robust and private machine learning models. Models designed to be robust to adversarial examples often overfit on training data to a larger extent than standard (non-robust) models. If a dataset contains private information, then any statistical test that separates training and test data by observing a model's outputs can represent a privacy breach, and if a model overfits on training data, these statistical tests become easier. In this work, we identify settings where standard models will overfit to a larger extent in comparison to robust models, and as empirically observed in previous works, settings where the opposite behavior occurs. Thus, it is not necessarily the case that privacy must be sacrificed to achieve robustness. The degree of overfitting naturally depends on the amount of data available for training. We go on to characterize how the training set size factors into the privacy risks exposed by training a robust model on a simple Gaussian data task, and show empirically that our findings hold on image classification benchmark datasets, such as CIFAR-10 and CIFAR-100.