The main objective of this work is to utilize state-of-the-art deep learning approaches for the identification of pulmonary embolism in CTPA-Scans for COVID-19 patients, provide an initial assessment of their performance and, ultimately, provide a fast-track prototype solution (system). We adopted and assessed some of the most popular convolutional neural network architectures through transfer learning approaches, to strive to combine good model accuracy with fast training. Additionally, we exploited one of the most popular one-stage object detection models for the localization (through object detection) of the pulmonary embolism regions-of-interests. The models of both approaches are trained on an original CTPA-Scan dataset, where we annotated of 673 CTPA-Scan images with 1,465 bounding boxes in total, highlighting pulmonary embolism regions-of-interests. We provide a brief assessment of some state-of-the-art image classification models by achieving validation accuracies of 91% in pulmonary embolism classification. Additionally, we achieved a precision of about 68% on average in the object detection model for the pulmonary embolism localization under 50% IoU threshold. For both approaches, we provide the entire training pipelines for future studies (step by step processes through source code). In this study, we present some of the most accurate and fast deep learning models for pulmonary embolism identification in CTPA-Scans images, through classification and localization (object detection) approaches for patients infected by COVID-19. We provide a fast-track solution (system) for the research community of the area, which combines both classification and object detection models for improving the precision of identifying pulmonary embolisms.
We present FedScale, a federated learning (FL) benchmarking suite with realistic datasets and a scalable runtime to enable reproducible FL research. FedScale datasets encompass a wide range of critical FL tasks, ranging from image classification and object detection to language modeling and speech recognition. Each dataset comes with a unified evaluation protocol using real-world data splits and evaluation metrics. To reproduce realistic FL behavior, FedScale contains a scalable and extensible runtime. It provides high-level APIs to implement FL algorithms, deploy them at scale across diverse hardware and software backends, and evaluate them at scale, all with minimal developer efforts. We combine the two to perform systematic benchmarking experiments and highlight potential opportunities for heterogeneity-aware co-optimizations in FL. FedScale is open-source and actively maintained by contributors from different institutions at http://fedscale.ai. We welcome feedback and contributions from the community.
The learning rate (LR) schedule is one of the most important hyper-parameters needing careful tuning in training DNNs. However, it is also one of the least automated parts of machine learning systems and usually costs significant manual effort and computing. Though there are pre-defined LR schedules and optimizers with adaptive LR, they introduce new hyperparameters that need to be tuned separately for different tasks/datasets. In this paper, we consider the question: Can we automatically tune the LR over the course of training without human involvement? We propose an efficient method, AutoLRS, which automatically optimizes the LR for each training stage by modeling training dynamics. AutoLRS aims to find an LR applied to every $\tau$ steps that minimizes the resulted validation loss. We solve this black-box optimization on the fly by Bayesian optimization (BO). However, collecting training instances for BO requires a system to evaluate each LR queried by BO's acquisition function for $\tau$ steps, which is prohibitively expensive in practice. Instead, we apply each candidate LR for only $\tau'\ll\tau$ steps and train an exponential model to predict the validation loss after $\tau$ steps. This mutual-training process between BO and the loss-prediction model allows us to limit the training steps invested in the BO search. We demonstrate the advantages and the generality of AutoLRS through extensive experiments of training DNNs for tasks from diverse domains using different optimizers. The LR schedules auto-generated by AutoLRS lead to a speedup of $1.22\times$, $1.43\times$, and $1.5\times$ when training ResNet-50, Transformer, and BERT, respectively, compared to the LR schedules in their original papers, and an average speedup of $1.31\times$ over state-of-the-art heavily-tuned LR schedules.
We propose an efficient and straightforward method for compressing deep convolutional neural networks (CNNs) that uses basis filters to represent the convolutional layers, and optimizes the performance of the compressed network directly in the basis space. Specifically, any spatial convolution layer of the CNN can be replaced by two successive convolution layers: the first is a set of three-dimensional orthonormal basis filters, followed by a layer of one-dimensional filters that represents the original spatial filters in the basis space. We jointly fine-tune both the basis and the filter representation to directly mitigate any performance loss due to the truncation. Generality of the proposed approach is demonstrated by applying it to several well known deep CNN architectures and data sets for image classification and object detection. We also present the execution time and power usage at different compression levels on the Xavier Jetson AGX processor.
Large scale image classification datasets often contain noisy labels. We take a principled probabilistic approach to modelling input-dependent, also known as heteroscedastic, label noise in these datasets. We place a multivariate Normal distributed latent variable on the final hidden layer of a neural network classifier. The covariance matrix of this latent variable, models the aleatoric uncertainty due to label noise. We demonstrate that the learned covariance structure captures known sources of label noise between semantically similar and co-occurring classes. Compared to standard neural network training and other baselines, we show significantly improved accuracy on Imagenet ILSVRC 2012 79.3% (+2.6%), Imagenet-21k 47.0% (+1.1%) and JFT 64.7% (+1.6%). We set a new state-of-the-art result on WebVision 1.0 with 76.6% top-1 accuracy. These datasets range from over 1M to over 300M training examples and from 1k classes to more than 21k classes. Our method is simple to use, and we provide an implementation that is a drop-in replacement for the final fully-connected layer in a deep classifier.
Deep convolutional neural networks accuracy is heavily impacted by rotations of the input data. In this paper, we propose a convolutional predictor that is invariant to rotations in the input. This architecture is capable of predicting the angular orientation without angle-annotated data. Furthermore, the predictor maps continuously the random rotation of the input to a circular space of the prediction. For this purpose, we use the roto-translation properties existing in the Scattering Transform Networks with a series of 3D Convolutions. We validate the results by training with upright and randomly rotated samples. This allows further applications of this work on fields like automatic re-orientation of randomly oriented datasets.
Within the context of autonomous driving, safety-related metrics for deep neural networks have been widely studied for image classification and object detection. In this paper, we further consider safety-aware correctness and robustness metrics specialized for semantic segmentation. The novelty of our proposal is to move beyond pixel-level metrics: Given two images with each having N pixels being class-flipped, the designed metrics should, depending on the clustering of pixels being class-flipped or the location of occurrence, reflect a different level of safety criticality. The result evaluated on an autonomous driving dataset demonstrates the validity and practicality of our proposed methodology.
Deep learning (DL) has achieved unprecedented success in a variety of tasks. However, DL systems are notoriously difficult to test and debug due to the lack of explainability of DL models and the huge test input space to cover. Generally speaking, it is relatively easy to collect a massive amount of test data, but the labeling cost can be quite high. Consequently, it is essential to conduct test selection and label only those selected "high quality" bug-revealing test inputs for test cost reduction. In this paper, we propose a novel test prioritization technique that brings order into the unlabeled test instances according to their bug-revealing capabilities, namely TestRank. Different from existing solutions, TestRank leverages both intrinsic attributes and contextual attributes of test instances when prioritizing them. To be specific, we first build a similarity graph on test instances and training samples, and we conduct graph-based semi-supervised learning to extract contextual features. Then, for a particular test instance, the contextual features extracted from the graph neural network (GNN) and the intrinsic features obtained with the DL model itself are combined to predict its bug-revealing probability. Finally, TestRank prioritizes unlabeled test instances in descending order of the above probability value. We evaluate the performance of TestRank on a variety of image classification datasets. Experimental results show that the debugging efficiency of our method significantly outperforms existing test prioritization techniques.
Deep Neural Networks (DNNs) have been shown vulnerable to Test-Time Evasion attacks (TTEs, or adversarial examples), which, by making small changes to the input, alter the DNN's decision. We propose an unsupervised attack detector on DNN classifiers based on class-conditional Generative Adversarial Networks (GANs). We model the distribution of clean data conditioned on the predicted class label by an Auxiliary Classifier GAN (AC-GAN). Given a test sample and its predicted class, three detection statistics are calculated based on the AC-GAN Generator and Discriminator. Experiments on image classification datasets under various TTE attacks show that our method outperforms previous detection methods. We also investigate the effectiveness of anomaly detection using different DNN layers (input features or internal-layer features) and demonstrate, as one might expect, that anomalies are harder to detect using features closer to the DNN's output layer.
Image classification methods are usually trained to perform predictions taking into account a predefined group of known classes. Real-world problems, however, may not allow for a full knowledge of the input and label spaces, making failures in recognition a hazard to deep visual learning. Open set recognition methods are characterized by the ability to correctly identify inputs of known and unknown classes. In this context, we propose GeMOS: simple and plug-and-play open set recognition modules that can be attached to pretrained Deep Neural Networks for visual recognition. The GeMOS framework pairs pre-trained Convolutional Neural Networks with generative models for open set recognition to extract open set scores for each sample, allowing for failure recognition in object recognition tasks. We conduct a thorough evaluation of the proposed method in comparison with state-of-the-art open set algorithms, finding that GeMOS either outperforms or is statistically indistinguishable from more complex and costly models.
Single-Image Super Resolution (SISR) is a classical computer vision problem and it has been studied for over decades. With the recent success of deep learning methods, recent work on SISR focuses solutions with deep learning methodologies and achieves state-of-the-art results. However most of the state-of-the-art SISR methods contain millions of parameters and layers, which limits their practical applications. In this paper, we propose a hardware (Synaptics Dolphin NPU) limitation aware, extremely lightweight quantization robust real-time super resolution network (XLSR). The proposed model's building block is inspired from root modules for Image classification. We successfully applied root modules to SISR problem, further more to make the model uint8 quantization robust we used Clipped ReLU at the last layer of the network and achieved great balance between reconstruction quality and runtime. Furthermore, although the proposed network contains 30x fewer parameters than VDSR its performance surpasses it on Div2K validation set. The network proved itself by winning Mobile AI 2021 Real-Time Single Image Super Resolution Challenge.
There has been a widely held view that visual representations (e.g., photographs and illustrations) do not depict negation, for example, one that can be expressed by a sentence "the train is not coming". This view is empirically challenged by analyzing the real-world visual representations of comic (manga) illustrations. In the experiment using image captioning tasks, we gave people comic illustrations and asked them to explain what they could read from them. The collected data showed that some comic illustrations could depict negation without any aid of sequences (multiple panels) or conventional devices (special symbols). This type of comic illustrations was subjected to further experiments, classifying images into those containing negation and those not containing negation. While this image classification was easy for humans, it was difficult for data-driven machines, i.e., deep learning models (CNN), to achieve the same high performance. Given the findings, we argue that some comic illustrations evoke background knowledge and thus can depict negation with purely visual elements.
Recently, collaborative learning proposed by Song and Chai has achieved remarkable improvements in image classification tasks by simultaneously training multiple classifier heads. However, huge memory footprints required by such multi-head structures may hinder the training of large-capacity baseline models. The natural question is how to achieve collaborative learning within a single network without duplicating any modules. In this paper, we propose four ways of collaborative learning among different parts of a single network with negligible engineering efforts. To improve the robustness of the network, we leverage the consistency of the output layer and intermediate layers for training under the collaborative learning framework. Besides, the similarity of intermediate representation and convolution kernel is also introduced to reduce the reduce redundant in a neural network. Compared to the method of Song and Chai, our framework further considers the collaboration inside a single model and takes smaller overhead. Extensive experiments on Cifar-10, Cifar-100, ImageNet32 and STL-10 corroborate the effectiveness of these four ways separately while combining them leads to further improvements. In particular, test errors on the STL-10 dataset are decreased by $9.28\%$ and $5.45\%$ for ResNet-18 and VGG-16 respectively. Moreover, our method is proven to be robust to label noise with experiments on Cifar-10 dataset. For example, our method has $3.53\%$ higher performance under $50\%$ noise ratio setting.
Superpixels are higher-order perceptual groups of pixels in an image, often carrying much more information than the raw pixels. There is an inherent relational structure to the relationship among different superpixels of an image such as adjacent superpixels are neighbours of each other. Our interest here is to treat these relative positions of various superpixels as relational information of an image. This relational information can convey higher-order spatial information about the image, such as the relationship between superpixels representing two eyes in an image of a cat. That is, two eyes are placed adjacent to each other in a straight line or the mouth is below the nose. Our motive in this paper is to assist computer vision models, specifically those based on Deep Neural Networks (DNNs), by incorporating this higher-order information from superpixels. We construct a hybrid model that leverages (a) Convolutional Neural Network (CNN) to deal with spatial information in an image and (b) Graph Neural Network (GNN) to deal with relational superpixel information in the image. The proposed model is learned using a generic hybrid loss function. Our experiments are extensive, and we evaluate the predictive performance of our proposed hybrid vision model on seven different image classification datasets from a variety of domains such as digit and object recognition, biometrics, medical imaging. The results demonstrate that the relational superpixel information processed by a GNN can improve the performance of a standard CNN-based vision system.
Radiologists usually observe anatomical regions of chest X-ray images as well as the overall image before making a decision. However, most existing deep learning models only look at the entire X-ray image for classification, failing to utilize important anatomical information. In this paper, we propose a novel multi-label chest X-ray classification model that accurately classifies the image finding and also localizes the findings to their correct anatomical regions. Specifically, our model consists of two modules, the detection module and the anatomical dependency module. The latter utilizes graph convolutional networks, which enable our model to learn not only the label dependency but also the relationship between the anatomical regions in the chest X-ray. We further utilize a method to efficiently create an adjacency matrix for the anatomical regions using the correlation of the label across the different regions. Detailed experiments and analysis of our results show the effectiveness of our method when compared to the current state-of-the-art multi-label chest X-ray image classification methods while also providing accurate location information.
There is an increasing interest in emulating Spiking Neural Networks (SNNs) on neuromorphic computing devices due to their low energy consumption. Recent advances have allowed training SNNs to a point where they start to compete with traditional Artificial Neural Networks (ANNs) in terms of accuracy, while at the same time being energy efficient when run on neuromorphic hardware. However, the process of training SNNs is still based on dense tensor operations originally developed for ANNs which do not leverage the spatiotemporally sparse nature of SNNs. We present here the first sparse SNN backpropagation algorithm which achieves the same or better accuracy as current state of the art methods while being significantly faster and more memory efficient. We show the effectiveness of our method on real datasets of varying complexity (Fashion-MNIST, Neuromophic-MNIST and Spiking Heidelberg Digits) achieving a speedup in the backward pass of up to 150x, and 85% more memory efficient, without losing accuracy.
Communication overhead hinders the scalability of large-scale distributed training. Gossip SGD, where each node averages only with its neighbors, is more communication-efficient than the prevalent parallel SGD. However, its convergence rate is reversely proportional to quantity $1-\beta$ which measures the network connectivity. On large and sparse networks where $1-\beta \to 0$, Gossip SGD requires more iterations to converge, which offsets against its communication benefit. This paper introduces Gossip-PGA, which adds Periodic Global Averaging into Gossip SGD. Its transient stage, i.e., the iterations required to reach asymptotic linear speedup stage, improves from $\Omega(\beta^4 n^3/(1-\beta)^4)$ to $\Omega(\beta^4 n^3 H^4)$ for non-convex problems. The influence of network topology in Gossip-PGA can be controlled by the averaging period $H$. Its transient-stage complexity is also superior to Local SGD which has order $\Omega(n^3 H^4)$. Empirical results of large-scale training on image classification (ResNet50) and language modeling (BERT) validate our theoretical findings.
In the paper of ExquisiteNetV1, the ability of classification of ExquisiteNetV1 is worse than DenseNet. In this article, we propose a faster and better model ExquisiteNetV2. We conduct many experiments to evaluate its performance. We test ExquisiteNetV2, ExquisiteNetV1 and other 9 well-known models on 15 credible datasets under the same condition. According to the experimental results, ExquisiteNetV2 gets the highest classification accuracy over half of the datasets. Important of all, ExquisiteNetV2 has fewest amounts of parameters. Besides, in most instances, ExquisiteNetV2 has fastest computing speed.
Federated Learning (FL) is a framework which enables distributed model training using a large corpus of decentralized training data. Existing methods aggregate models disregarding their internal representations, which are crucial for training models in vision tasks. System and statistical heterogeneity (e.g., highly imbalanced and non-i.i.d. data) further harm model training. To this end, we introduce a method, called FedProto, which computes client deviations using margins of prototypical representations learned on distributed data, and applies them to drive federated optimization via an attention mechanism. In addition, we propose three methods to analyse statistical properties of feature representations learned in FL, in order to elucidate the relationship between accuracy, margins and feature discrepancy of FL models. In experimental analyses, FedProto demonstrates state-of-the-art accuracy and convergence rate across image classification and semantic segmentation benchmarks by enabling maximum margin training of FL models. Moreover, FedProto reduces uncertainty of predictions of FL models compared to the baseline. To our knowledge, this is the first work evaluating FL models in dense prediction tasks, such as semantic segmentation.
Knowledge distillation (KD), transferring knowledge from a cumbersome teacher model to a lightweight student model, has been investigated to design efficient neural architectures. Generally, the objective function of KD is the Kullback-Leibler (KL) divergence loss between the softened probability distributions of the teacher model and the student model with the temperature scaling hyperparameter tau. Despite its widespread use, few studies have discussed the influence of such softening on generalization. Here, we theoretically show that the KL divergence loss focuses on the logit matching when tau increases and the label matching when tau goes to 0 and empirically show that the logit matching is positively correlated to performance improvement in general. From this observation, we consider an intuitive KD loss function, the mean squared error (MSE) between the logit vectors, so that the student model can directly learn the logit of the teacher model. The MSE loss outperforms the KL divergence loss, explained by the difference in the penultimate layer representations between the two losses. Furthermore, we show that sequential distillation can improve performance and that KD, particularly when using the KL divergence loss with small tau, mitigates the label noise. The code to reproduce the experiments is publicly available online at https://github.com/jhoon-oh/kd_data/.