Due to the black-box nature of deep learning models, there is a recent development of solutions for visual explanations of CNNs. Given the high cost of user studies, metrics are necessary to compare and evaluate these different methods. In this paper, we critically analyze the Deletion Area Under Curve (DAUC) and Insertion Area Under Curve (IAUC) metrics proposed by Petsiuk et al. (2018). These metrics were designed to evaluate the faithfulness of saliency maps generated by generic methods such as Grad-CAM or RISE. First, we show that the actual saliency score values given by the saliency map are ignored as only the ranking of the scores is taken into account. This shows that these metrics are insufficient by themselves, as the visual appearance of a saliency map can change significantly without the ranking of the scores being modified. Secondly, we argue that during the computation of DAUC and IAUC, the model is presented with images that are out of the training distribution which might lead to an unreliable behavior of the model being explained. To complement DAUC/IAUC, we propose new metrics that quantify the sparsity and the calibration of explanation methods, two previously unstudied properties. Finally, we give general remarks about the metrics studied in this paper and discuss how to evaluate them in a user study.
Hubei University of Technology（湖北工业大学）；Central China Normal University（华中师范大学）
The black-box nature of Deep Neural Networks (DNNs) severely hinders its performance improvement and application in specific scenes. In recent years, class activation mapping-based method has been widely used to interpret the internal decisions of models in computer vision tasks. However, when this method uses backpropagation to obtain gradients, it will cause noise in the saliency map, and even locate features that are irrelevant to decisions. In this paper, we propose an Absolute value Class Activation Mapping-based (Abs-CAM) method, which optimizes the gradients derived from the backpropaga…
Wroclaw University of Science and Technology（弗罗茨瓦夫理工大学）
In this paper, an enhancement technique for the class activation mapping methods such as gradient-weighted class activation maps or excitation backpropagation is proposed to present the visual explanations of decisions from convolutional neural network-based models. The proposed idea, called Gradual Extrapolation, can supplement any method that generates a heatmap picture by sharpening the output. Instead of producing a coarse localization map that highlights the important predictive regions in the image, the proposed method outputs the specific shape that most contributes to the model output.…
Bosch Center for Artificial Intelligence（博世人工智能中心）
Recently, there has been a growing interest in developing saliency methods that provide visual explanations of network predictions. Still, the usability of existing methods is limited to image classification models. To overcome this limitation, we extend the existing approaches to generate grid saliencies, which provide spatially coherent visual explanations for (pixel-level) dense prediction networks. As the proposed grid saliency allows to spatially disentangle the object and its context, we specifically explore its potential to produce context explanations for semantic segmentation networks…
Deep learning is widely used to decode the electroencephalogram (EEG) signal. However, there are few attempts to specifically investigate how to explain the EEG-based deep learning models. We conduct a review to summarize the existing works explaining the EEG-based deep learning model. Unfortunately, we find that there is no appropriate method to explain them. Based on the characteristic of EEG data, we suggest a context-aware perturbation method to generate a saliency map from the perspective of the raw EEG signal. Moreover, we also justify that the context information can be used to suppress…
Explaining a classification result produced by an image- and video-classification model is one of the important but challenging issues in computer vision. Many methods have been proposed for producing heat-map--based explanations for this purpose, including ones based on the white-box approach that uses the internal information of a model (e.g., LRP, Grad-CAM, and Grad-CAM++) and ones based on the black-box approach that does not use any internal information (e.g., LIME, SHAP, and RISE). We propose a new black-box method BOREx (Bayesian Optimization for Refinement of visual model Explanation) …
Leibniz Universität Hannover（汉诺威莱布尼茨大学）；dSPACE GmbH（dSPACE有限公司）
In recent years, an abundance of feature attribution methods for explaining neural networks have been developed. Especially in the field of computer vision, many methods for generating saliency maps providing pixel attributions exist. However, their explanations often contradict each other and it is not clear which explanation to trust. A natural solution to this problem is the aggregation of multiple explanations. We present and compare different pixel-based aggregation schemes with the goal of generating a new explanation, whose fidelity to the model's decision is higher than each individual…
Boston University（波士顿大学）；Adobe Research（Adobe研究院）；Adobe Document Cloud（Adobe文档云）；University of Virginia（弗吉尼亚大学）；MIT-IBM Watson AI Lab（麻省理工学院-IBM沃森人工智能实验室）
We propose D-RISE, a method for generating visual explanations for the predictions of object detectors. Utilizing the proposed similarity metric that accounts for both localization and categorization aspects of object detection allows our method to produce saliency maps that show image areas that most affect the prediction. D-RISE can be considered "black-box" in the software testing sense, as it only needs access to the inputs and outputs of an object detector. Compared to gradient-based methods, D-RISE is more general and agnostic to the particular type of object detector being tested, and d…
Simon Fraser University（西蒙弗雷泽大学）；The University of British Columbia（不列颠哥伦比亚大学）
Being able to explain the prediction to clinical end-users is a necessity to leverage the power of AI models for clinical decision support. For medical images, saliency maps are the most common form of explanation. The maps highlight important features for AI model's prediction. Although many saliency map methods have been proposed, it is unknown how well they perform on explaining decisions on multi-modal medical images, where each modality/channel carries distinct clinical meanings of the same underlying biomedical phenomenon. Understanding such modality-dependent features is essential for c…
University of British Columbia（不列颠哥伦比亚大学）；Huawei（华为）
The opacity of deep learning models constrains their debugging and improvement. Augmenting deep models with saliency-based strategies, such as attention, has been claimed to help get a better understanding of the decision-making process of black-box models. However, some recent works challenged saliency's faithfulness in the field of Natural Language Processing (NLP), questioning attention weights' adherence to the true decision-making process of the model. We add to this discussion by evaluating the faithfulness of in-model saliency applied to a video processing task for the first time, namel…
IIT Kharagpur（印度理工学院卡拉格普尔分校）；University of Poitiers（普瓦捷大学）；IIT Jodhpur（印度理工学院焦特布尔分校）
Deep neural networks have shown their profound impact on achieving human level performance in visual saliency prediction. However, it is still unclear how they learn the task and what it means in terms of understanding human visual system. In this work, we develop a technique to derive explainable saliency models from their corresponding deep neural architecture based saliency models by applying human perception theories and the conventional concepts of saliency. This technique helps us understand the learning pattern of the deep network at its intermediate layers through their activation maps…
In this paper, we propose an efficient saliency map generation method, called Group score-weighted Class Activation Mapping (Group-CAM), which adopts the "split-transform-merge" strategy to generate saliency maps. Specifically, for an input image, the class activations are firstly split into groups. In each group, the sub-activations are summed and de-noised as an initial mask. After that, the initial masks are transformed with meaningful perturbations and then applied to preserve sub-pixels of the input (i.e., masked inputs), which are then fed into the network to calculate the confidence sco…
Beihang University（北京航空航天大学）；University of Central Florida（中佛罗里达大学）
In the area of human fixation prediction, dozens of computational saliency models are proposed to reveal certain saliency characteristics under different assumptions and definitions. As a result, saliency model benchmarking often requires several evaluation metrics to simultaneously assess saliency models from multiple perspectives. However, most computational metrics are not designed to directly measure the perceptual similarity of saliency maps so that the evaluation results may be sometimes inconsistent with the subjective impression. To address this problem, this paper first conducts exten…
Massachusetts Institute of Technology（麻省理工学院）；Google（谷歌）
How best to evaluate a saliency model's ability to predict where humans look in images is an open research question. The choice of evaluation metric depends on how saliency is defined and how the ground truth is represented. Metrics differ in how they rank saliency models, and this results from how false positives and false negatives are treated, whether viewing biases are accounted for, whether spatial deviations are factored in, and how the saliency maps are pre-processed. In this paper, we provide an analysis of 8 different evaluation metrics and their properties. With the help of systemati…
UCSF（加州大学旧金山分校）；SFVAMC（旧金山退伍军人事务医疗中心）；Helen Diller Family Comprehensive Cancer Center（海伦·迪勒家庭综合癌症中心）；Institute for Neurodegenerative Diseases（神经退行性疾病研究所）；Bakar Computational Health Sciences Institute（巴卡尔计算健康科学研究所）
In high-stakes applications of machine learning models, interpretability methods provide guarantees that models are right for the right reasons. In medical imaging, saliency maps have become the standard tool for determining whether a neural model has learned relevant robust features, rather than artefactual noise. However, saliency maps are limited to local model explanation because they interpret predictions on an image-by-image basis. We propose aggregating saliency globally, using semantic segmentation masks, to provide quantitative measures of model bias across a dataset. To evaluate glob…
Channel pruning is used to reduce the number of weights in a Convolutional Neural Network (CNN). Channel pruning removes slices of the weight tensor so that the convolution layer remains dense. The removal of these weight slices from a single layer causes mismatching number of feature maps between layers of the network. A simple solution is to force the number of feature map between layers to match through the removal of weight slices from subsequent layers. This additional constraint becomes more apparent in DNNs with branches where multiple channels need to be pruned together to keep the net…
Utrecht University（乌得勒支大学）；University of Essex（埃塞克斯大学）；London South Bank University（伦敦南岸大学）；Noldus Information Technology（诺尔达斯信息技术公司）
Deep learning approaches have been established as the main methodology for video classification and recognition. Recently, 3-dimensional convolutions have been used to achieve state-of-the-art performance in many challenging video datasets. Because of the high level of complexity of these methods, as the convolution operations are also extended to additional dimension in order to extract features from them as well, providing a visualization for the signals that the network interpret as informative, is a challenging task. An effective notion of understanding the network's inner-workings would b…
Saliency detection has been widely studied because it plays an important role in various vision applications, but it is difficult to evaluate saliency systems because each measure has its own bias. In this paper, we first revisit the problem of applying the widely used saliency metrics on modern Convolutional Neural Networks(CNNs). Our investigation shows the saliency datasets have been built based on different choices of parameters and CNNs are designed to fit a dataset-specific distribution. Secondly, we show that the Shuffled Area Under Curve(S-AUC) metric still suffers from spatial biases.…
Werner Reichardt Centre for Integrative Neuroscience, University of Tübingen（图宾根大学维尔纳·赖卡特综合神经科学中心）
Dozens of new models on fixation prediction are published every year and compared on open benchmarks such as MIT300 and LSUN. However, progress in the field can be difficult to judge because models are compared using a variety of inconsistent metrics. Here we show that no single saliency map can perform well under all metrics. Instead, we propose a principled approach to solve the benchmarking problem by separating the notions of saliency models, maps and metrics. Inspired by Bayesian decision theory, we define a saliency model to be a probabilistic model of fixation density prediction and a s…
With the increased focus on visual attention (VA) in the last decade, a large number of computational visual saliency methods have been developed over the past few years. These models are traditionally evaluated by using performance evaluation metrics that quantify the match between predicted saliency and fixation data obtained from eye-tracking experiments on human observers. Though a considerable number of such metrics have been proposed in the literature, there are notable problems in them. In this work, we discuss shortcomings in existing metrics through illustrative examples and propose a…
Saliency strikes back: How filtering out high frequencies improves white-box explanations
Carney Institute for Brain Science, Brown University（布朗大学卡尼脑科学研究所）；CerCo, CNRS（法国国家科学研究中心CerCo研究所）；SNCF（法国国家铁路公司）；Institute of Mathematics of Toulouse, University Paul Sabatier（保罗萨巴蒂埃大学图卢兹数学研究所）
IBM Research（IBM研究院）；Cardiff University（卡迪夫大学）；Crime and Security Research Institute（犯罪与安全研究所）；Booz Allen Hamilton（博思艾伦咨询公司）；CCDC Army Research Laboratory（CCDC陆军研究实验室）
Saliency maps are a popular approach to creating post-hoc explanations of image classifier outputs. These methods produce estimates of the relevance of each pixel to the classification output score, which can be displayed as a saliency map that highlights important pixels. Despite a proliferation of such methods, little effort has been made to quantify how good these saliency maps are at capturing the true relevance of the pixels to the classifier output (i.e. their "fidelity"). We therefore investigate existing metrics for evaluating the fidelity of saliency methods (i.e. saliency metrics). W…
National University of Defense Technology（国防科技大学）；National University of Singapore（新加坡国立大学）
Salient object detection has seen remarkable progress driven by deep learning techniques. However, most of deep learning based salient object detection methods are black-box in nature and lacking in interpretability. This paper proposes the first self-explanatory saliency detection network that explicitly exploits low- and high-level features for salient object detection. We demonstrate that such supportive clues not only significantly enhances performance of salient object detection but also gives better justified detection results. More specifically, we develop a multi-stage saliency encoder…
The computation and memory needed for Convolutional Neural Network (CNN) inference can be reduced by pruning weights from the trained network. Pruning is guided by a pruning saliency, which heuristically approximates the change in the loss function associated with the removal of specific weights. Many pruning signals have been proposed, but the performance of each heuristic depends on the particular trained network. This leaves the data scientist with a difficult choice. When using any one saliency metric for the entire pruning process, we run the risk of the metric assumptions being invalidat…
Signature Activation: A Sparse Signal View for Holistic Saliency
Harvard University（哈佛大学）；Broad Institute of Massachusetts Institute of Technology and Harvard（麻省理工学院与哈佛大学博德研究所）；Massachusetts General Hospital（麻省总医院）；Harvard Medical School（哈佛医学院）
Classification Metrics for Image Explanations: Towards Building Reliable XAI-Evaluations
Fraunhofer Institute for Manufacturing Engineering and Automation IPA（弗劳恩霍夫制造工程与自动化研究所）；Institute of Industrial Manufacturing and Management IFF, University of Stuttgart（斯图加特大学工业制造与管理研究所）
TU Berlin（柏林工业大学）；Fraunhofer HHI（弗劳恩霍夫海因里希·赫兹研究所）；Leibniz Institute of Agricultural Engineering and Bioeconomy e.V. (ATB)（莱布尼茨农业工程与生物经济研究所）；BIFOLD – Berlin Institute for the Foundations of Learning and Data（柏林学习与数据基础研究所）；University of Potsdam（波茨坦大学）
University of Florence（佛罗伦萨大学）；University of Siena（锡耶纳大学）
In the last three decades, human visual attention has been a topic of great interest in various disciplines. In computer vision, many models have been proposed to predict the distribution of human fixations on a visual stimulus. Recently, thanks to the creation of large collections of data, machine learning algorithms have obtained state-of-the-art performance on the task of saliency map estimation. On the other hand, computational models of scanpath are much less studied. Works are often only descriptive or task specific. This is due to the fact that the scanpath is harder to model because it…
Xiamen University（厦门大学）；Australian National University（澳大利亚国立大学）；Peng Cheng Laboratory（鹏城实验室）；Zhejiang University（浙江大学）
The explanation for deep neural networks has drawn extensive attention in the deep learning community over the past few years. In this work, we study the visual saliency, a.k.a. visual explanation, to interpret convolutional neural networks. Compared to iteration based saliency methods, single backward pass based saliency methods benefit from faster speed, and they are widely used in downstream visual tasks. Thus, we focus on single backward pass based methods. However, existing methods in this category struggle to uccessfully produce fine-grained saliency maps concentrating on specific target…
Saliency map estimation in computer vision aims to estimate the locations where people gaze in images. Since people tend to look at objects in images, the parameters of the model pretrained on ImageNet for image classification are useful for the saliency map estimation. However, there is no research on the relationship between the image classification accuracy and the performance of the saliency map estimation. In this paper, it is shown that there is a strong correlation between image classification accuracy and saliency map estimation accuracy. We also investigated the effective architecture…
Saliency methods generating visual explanatory maps representing the importance of image pixels for model classification is a popular technique for explaining neural network decisions. Hierarchical dynamic masks (HDM), a novel explanatory maps generation method, is proposed in this paper to enhance the granularity and comprehensiveness of saliency maps. First, we suggest the dynamic masks (DM), which enables multiple small-sized benchmark mask vectors to roughly learn the critical information in the image through an optimization method. Then the benchmark mask vectors guide the learning of lar…
Instance-level quantitative saliency in multiple sclerosis lesion segmentation
Translational Imaging in Neurology (ThINk)（神经学转化成像（ThINk））；Department of Medicine and Biomedical Engineering, University Hospital Basel and University of Basel（医学与生物医学工程系，巴塞尔大学医院和巴塞尔大学）；Department of Neurology, University Hospital Basel（神经病学系，巴塞尔大学医院）；Research Center for Clinical Neuroimmunology and Neuroscience Basel (RC2NB)（巴塞尔临床神经免疫学和神经科学研究中心（RC2NB））；MedGIFT, Institute of Informatics, School of Management, HES-SO Valais-Wallis University of Applied Sciences and Arts Western Switzerland（MedGIFT，信息学院，管理学院，瓦莱-迪-瓦尔省应用科学与艺术西方瑞士大学）；CIBM Center for Biomedical Imaging, Lausanne, Switzerland（洛桑瑞士生物医学成像中心）；Radiology Department, Lausanne University Hospital (CHUV) and University of Lausanne（放射科，洛桑大学医院（CHUV）和洛桑大学）
Massachusetts General Hospital（麻省总医院）；Harvard Medical School（哈佛医学院）；Shiv Nadar University（希夫纳达尔大学）；Georgia Institute of Technology（佐治亚理工学院）；Massachusetts Institute of Technology（麻省理工学院）
Saliency maps have become a widely used method to make deep learning models more interpretable by providing post-hoc explanations of classifiers through identification of the most pertinent areas of the input medical image. They are increasingly being used in medical imaging to provide clinically plausible explanations for the decisions the neural network makes. However, the utility and robustness of these visualization maps has not yet been rigorously examined in the context of medical imaging. We posit that trustworthiness in this context requires 1) localization utility, 2) sensitivity to m…
Recently, many methods to interpret and visualize deep neural network predictions have been proposed and significant progress has been made. However, a more class-discriminative and visually pleasing explanation is required. Thus, this paper proposes a region-based approach that estimates feature importance in terms of appropriately segmented regions. By fusing the saliency maps generated from multi-scale segmentations, a more class-discriminative and visually pleasing map is obtained. We incorporate this regional multi-scale concept into a prediction difference method that is model-agnostic. …
Symbolic techniques based on Satisfiability Modulo Theory (SMT) solvers have been proposed for analyzing and verifying neural network properties, but their usage has been fairly limited owing to their poor scalability with larger networks. In this work, we propose a technique for combining gradient-based methods with symbolic techniques to scale such analyses and demonstrate its application for model explanation. In particular, we apply this technique to identify minimal regions in an input that are most relevant for a neural network's prediction. Our approach uses gradient information (based …
A deep feature based saliency model (DeepFeat) is developed to leverage the understanding of the prediction of human fixations. Traditional saliency models often predict the human visual attention relying on few level image cues. Although such models predict fixations on a variety of image complexities, their approaches are limited to the incorporated features. In this study, we aim to provide an intuitive interpretation of convolu- tional neural network deep features by combining low and high level visual factors. We exploit four evaluation metrics to evaluate the correspondence between the p…
Point cloud quality assessment (PCQA) has become an appealing research field in recent days. Considering the importance of saliency detection in quality assessment, we propose an effective full-reference PCQA metric which makes the first attempt to utilize the saliency information to facilitate quality prediction, called point cloud quality assessment using 3D saliency maps (PQSM). Specifically, we first propose a projection-based point cloud saliency map generation method, in which depth information is introduced to better reflect the geometric characteristics of point clouds. Then, we constr…
