Semantic Bias Research Articles

Effects of Individual Differences and Prosodic Focus on the Interpretation of Quantity Scalar Terms in Mandarin-Speaking 3- to 8-Year-Olds.

This study focuses on examining how individual differences, including biological, linguistic, and cognitive traits, and prosodic focus affect the computation biases and reaction time (RT) associated with quantity scalar terms in Mandarin-speaking children aged 3-8 years. The participants of this study were 27 Mandarin-speaking children aged 3-8 years. They completed a computer-based sentence evaluation task, and their receptive vocabulary, nonverbal IQ, and theory of mind (ToM) skills were assessed. Additionally, parents provided insights into their children's executive functions, including working memory, planning, regulation, and inhibition abilities, through a questionnaire reflecting daily performance. Mandarin-speaking 3- to 8-year-olds showed pervasive quantifier semantic biases versus bimodally distributed ad hoc semantic/pragmatic biases. Their quantifier pragmatic bias increased with age, working memory, and planning abilities but decreased with first-order ToM. In contrast, their ad hoc pragmatic bias improved with second-order ToM, working memory, planning, and inhibition abilities but decreased with age and receptive vocabulary. Prosodic focus reduced the number of hesitators and minimized the RT differences between hesitators and pragmatic/semantic responders. Children show a higher overall pragmatic bias in ad hoc compared to quantifier scalar terms, alongside notable individual differences. Quantifier and ad hoc scalar terms appear to have different initial interpretations, with the former leaning toward a semantic interpretation and the latter toward a pragmatic one. Prosodic focus reduced hesitation and encouraged further processing, although it did not significantly alter interpretation biases. Future studies should employ larger sample sizes and implicit measures to further explore these findings.

Impact of SVM-based Poisoning on the Semantic Recognition of Sounds

Abstract. Machine learning is a technique that enables computers to learn from data and make predictions or decisions, data poisoning is the process of machine learning training where malicious samples are put in to make the model predictions or classifications less accurate. Data poisoning attacks help to reveal security vulnerabilities in AI systems. In this paper, we study Support Vector Machine (SVM) poisoning for sound recognition techniques, using AISHELL-3 dataset training data, from which we find the most vulnerable features for SVM poisoning. In the field of speech recognition, SVM can be applied to speech feature extraction, speech classification and speech synthesis to find the best hyperplane by finding the maximum margin optimization for effective classification and recognition of speech signals. Experiments have resulted in biased semantic recognition of sounds, output of incorrect speech, reduced accuracy of model classification and generation of incorrect decision boundaries. The role of this research paper is to investigate whether SVM poisoning affects the semantic recognition of sounds, and the result of the research is that it does cause semantic bias.

Dual De-confounded Causal Intervention method for knowledge graph error detection

Due to the Knowledge Graph (KG) construction process, erroneous triples are virtually inevitable to be introduced into real-world KGs. Since these errors hinder the expressiveness and applicability of KGs, the development of knowledge graph error detection (KGED) methods is necessary. Despite the overall effectiveness of current KGED methods, their capacity to identify challenging errors is limited. In this work, we conduct empirical studies and find that previous works introduce structural and semantic bias, impeding the identification of erroneous triples, especially in challenging cases. To address this issue, we design a causal graph for the KGED task and propose a Dual De-confounded Causal Intervention (DuDCI) method for debiasing. Firstly, DuDCI utilizes the neighborhood and textual descriptions of triples to calculate their graph and text embeddings. Next, a Causal De-confounded Module is constructed to mitigate the impact of shortcuts caused by the bias through the front-door adjustment. Furthermore, we introduce Disentanglement Constraints to disentangle the information expressed by each embedding, thereby facilitating further bias mitigation. Experimental results on three widely used KGED datasets validate the effectiveness of DuDCI and demonstrate that DuDCI outperforms current KGED methods, with an improvement of at least 2.2%, especially in more challenging noise scenarios.

Uncertainty guided semi-supervised few-shot segmentation with prototype level fusion

Few-Shot Semantic Segmentation (FSS) aims to tackle the challenge of segmenting novel categories with limited annotated data. However, given the diversity among support-query pairs, transferring meta-knowledge to unseen categories poses a significant challenge, particularly in scenarios featuring substantial intra-class variance within an episode task. To alleviate this issue, we propose the Uncertainty Guided Adaptive Prototype Network (UGAPNet) for semi-supervised few-shot semantic segmentation. The key innovation lies in the generation of reliable pseudo-prototypes as an additional supplement to alleviate intra-class semantic bias. Specifically, we employ a shared meta-learner to produce segmentation results for unlabeled images in the pseudo-label prediction module. Subsequently, we incorporate an uncertainty estimation module to quantify the difference between prototypes extracted from query and support images, facilitating pseudo-label denoising. Utilizing these refined pseudo-label samples, we introduce a prototype rectification module to obtain effective pseudo-prototypes and generate a generalized adaptive prototype for the segmentation of query images. Furthermore, generalized few-shot semantic segmentation extends the paradigm of few-shot semantic segmentation by simultaneously segmenting both unseen and seen classes during evaluation. To address the challenge of confusion region prediction between these two categories, we further propose a novel Prototype-Level Fusion Strategy in the prototypical contrastive space. Extensive experiments conducted on two benchmarks demonstrate the effectiveness of the proposed UGAPNet and prototype-level fusion strategy. Our source code will be available on https://github.com/WHL182/UGAPNet.

Olfaction modulates cortical arousal independent of perceived odor intensity and pleasantness

Throughout history, various odors have been harnessed to invigorate or relax the mind. The mechanisms underlying odors’ diverse arousal effects remain poorly understood. We conducted five experiments (184 participants) to investigate this issue, using pupillometry, electroencephalography, and the attentional blink paradigm, which exemplifies the limit in attentional capacity. Results demonstrated that exposure to citral, compared to vanillin, enlarged pupil size, reduced resting-state alpha oscillations and alpha network efficiency, augmented beta-gamma oscillations, and enhanced the coordination between parietal alpha and frontal beta-gamma activities. In parallel, it attenuated the attentional blink effect. These effects were observed despite citral and vanillin being comparable in perceived odor intensity, pleasantness, and nasal pungency, and were unlikely driven by semantic biases. Our findings reveal that odors differentially alter the small-worldness of brain network architecture, and thereby brain state and arousal. Furthermore, they establish arousal as a unique dimension in olfactory space, distinct from intensity and pleasantness.

Developmental refinements to neural attentional state during semantic memory retrieval through adolescence

Despite the fact that attention undergoes protracted development, little is known about how it may support memory refinements in childhood and adolescence. Here, we asked whether people differentially focus their attention on semantic or perceptual information over development during memory retrieval. First, we trained a multivoxel classifier to characterize whole-brain neural patterns reflecting semantic versus perceptual attention in a cued attention task. We then used this classifier to quantify how attention varied in a separate dataset in which children, adolescents, and adults retrieved autobiographical, semantic, and episodic memories. All age groups demonstrated a semantic attentional bias during memory retrieval, with significant age differences in this bias during the semantic task. Trials began with a preparatory picture cue followed by a retrieval question, which allowed us to ask whether attentional biases varied by trial period. Adults showed a semantic bias earlier during the picture cues, whereas adolescents showed this bias during the question. Adults and adolescents also engaged different brain regions—superior parietal cortex and ventral visual regions, respectively—during preparatory picture cues. Our results demonstrate that retrieval-related attention undergoes refinement beyond childhood. These findings suggest that alongside expanding semantic knowledge, attention-related changes may support the maturation of factual knowledge retrieval.

Artificial intelligence in judicial adjudication: Semantic biasness classification and identification in legal judgement (SBCILJ)

History reveals that human societies have suffered in terms of social justice due to cognitive bias. Semantic bias tends to amplify cognitive bias. Therefore, the presence of cognitive biases in extensive historical data can potentially result in unethical and allegedly inhumane predictions since AI systems are trained on this data. The innovation of artificial intelligence and its rapid integration across disciplines has prompted questions regarding the subjectivity of the technology. Current research focuses the semantic bias in legal judgment to increase the legitimacy of training data. By the application of general-purpose Artificial Intelligence (AI) algorithms, we classify and detect the semantics bias that is present in the Chinese Artificial Intelligence and Law (CAIL) dataset. Our findings demonstrate that AI models acquire superior prediction power in the CAIL dataset, which is comprised of hundreds of cases, compared to a structured professional risk assessment tool. To assist legal practitioners during this process, innovative approaches that are based on AI may be implemented inside the legal arena. To accomplish this objective, we suggested a classification model for semantic bias that is related to the classification and identification of semantic biases in legal judgment. Our proposed model legal field uses the example of categorization along with the identification of the CAIL dataset. This will be accomplished by identifying the semantics biases in judicial decisions. We used different types of classifiers such as the Support Vector Machine (SVM), Naïve-Bayes (NB), Multi-Layer Perceptron (MLP), and the K-Nearest Neighbour (KNN) to come across the preferred results. SVM got 96.90 %, NB has 88.80 %, MLP has 86.75 % and KNN achieved 85.66 % accuracy whereas SVM achieved greater accuracy as compared to other models. Additionally, we demonstrate that we were able to get a relatively high classification performance when predicting outcomes based just on the semantic bias categorization in judicial judgments that determine the outcome of the case.

Quantum Interference Model for Semantic Biases of Glosses in Word Sense Disambiguation

Word Sense Disambiguation (WSD) aims to determine the meaning of the target word according to the given context. Currently, a single representation enhanced by glosses from different dictionaries or languages is used to characterize each word sense. By analyzing the similarity between glosses of the same word sense, we find semantic biases among them, revealing that the glosses have their own descriptive perspectives. Therefore, the traditional approach of integrating all glosses by a single representation results in failing to present the unique semantics revealed by the individual glosses. In this paper, a quantum superposition state is employed to formalize the representations of multiple glosses of the same word sense to reveal their distributions. Furthermore, the quantum interference model is leveraged to calculate the probability that the target word belongs to this superposition state. The advantage is that the interference term can be regarded as a confidence level to guide word sense recognition. Finally, experiments are performed under standard WSD evaluation framework and the latest cross-lingual datasets, and the results verify the effectiveness of our model.

MGSGA: Multi-grained and Semantic-Guided Alignment for Text-Video Retrieval

In the text-video retrieval task, the objective is to calculate the similarity between a text and a video, and rank the relevant candidates higher. Most existing methods only consider the text-video semantic alignment in the global view. But using mean-pooling to obtain global semantics and simply aligning text and video in the global view may lead to semantic bias. In addition, some methods utilize offline object detectors or sentence parsers to obtain entity-level information in text and video and achieve local alignment. However, inaccurate detection introduces possible errors and such approaches prevent models from being trained end-to-end for retrieval. To overcome these limitations, we propose multi-grained and semantic-guided alignment for text-video retrieval in this paper, which can achieve fine-grained alignment based on video frames and text words, local alignment based on semantic centers, and global alignment. Specially, we explore summary semantics of text and video to guide the local alignment based on semantic centers for we believe that the importance of each semantic center is determined by summary semantics. We evaluate our approach on four benchmark datasets of MSRVTT, MSVD, ActivityNet Captions, and DiDeMo, achieving better performance than most existing methods.

A novel transformer-based aggregation model for predicting gene mutations in lung adenocarcinoma.

In recent years, predicting gene mutations on whole slide imaging (WSI) has gained prominence. The primary challenge is extracting global information and achieving unbiased semantic aggregation. To address this challenge, we propose a novel Transformer-based aggregation model, employing a self-learning weight aggregation mechanism to mitigate semantic bias caused by the abundance of features in WSI. Additionally, we adopt a random patch training method, which enhances model learning richness by randomly extracting feature vectors from WSI, thus addressing the issue of limited data. To demonstrate the model's effectiveness in predicting gene mutations, we leverage the lung adenocarcinoma dataset from Shandong Provincial Hospital for prior knowledge learning. Subsequently, we assess TP53, CSMD3, LRP1B, and TTN gene mutations using lung adenocarcinoma tissue pathology images and clinical data from The Cancer Genome Atlas (TCGA). The results indicate a notable increase in the AUC (Area Under the ROC Curve) value, averaging 4%, attesting to the model's performance improvement. Our research offers an efficient model to explore the correlation between pathological image features and molecular characteristics in lung adenocarcinoma patients. This model introduces a novel approach to clinical genetic testing, expected to enhance the efficiency of identifying molecular features and genetic testing in lung adenocarcinoma patients, ultimately providing more accurate and reliable results for related studies.

Toward Egocentric Compositional Action Anticipation with Adaptive Semantic Debiasing

Predicting the unknown from the first-person perspective is expected as a necessary step toward machine intelligence, which is essential for practical applications including autonomous driving and robotics. As a human-level task, egocentric action anticipation aims at predicting an unknown action seconds before it is performed from the first-person viewpoint. Egocentric actions are usually provided as verb-noun pairs; however, predicting the unknown action may be trapped in insufficient training data for all possible combinations. Therefore, it is crucial for intelligent systems to use limited known verb-noun pairs to predict new combinations of actions that have never appeared, which is known as compositional generalization. In this article, we are the first to explore the egocentric compositional action anticipation problem, which is more in line with real-world settings but neglected by existing studies. Whereas prediction results are prone to suffer from semantic bias considering the distinct difference between training and test distributions, we further introduce a general and flexible adaptive semantic debiasing framework that is compatible with different deep neural networks. To capture and mitigate semantic bias, we can imagine one counterfactual situation where no visual representations have been observed and only semantic patterns of observation are used to predict the next action. Instead of the traditional counterfactual analysis scheme that reduces semantic bias in a mindless way, we devise a novel counterfactual analysis scheme to adaptively amplify or penalize the effect of semantic experience by considering the discrepancy both among categories and among examples. We also demonstrate that the traditional counterfactual analysis scheme is a special case of the devised adaptive counterfactual analysis scheme. We conduct experiments on three large-scale egocentric video datasets. Experimental results verify the superiority and effectiveness of our proposed solution.

Patch-Mixing Contrastive Regularization for Few-Label Semi-Supervised Learning

Recently, consistency regularization has become a fundamental component in semi-supervised learning, which tries to make the network's predictions on unlabeled data to be invariant to perturbations. However, its performance decreases drastically when there are scarce labels, e.g., 2 labels per category. In this paper, we analyze the semantic bias problem in consistency regularization for semi-supervised learning and find that this problem stems from imposing consistency regularization on some semantically biased positive sample pairs derived from indispensable data augmentation. Based on the above analysis, we propose a patch-mixing contrastive regularization approach called <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$p$</tex-math></inline-formula> -Mix, for semi-supervised learning with scarce labels. In <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$p$</tex-math></inline-formula> -Mix, the magnitude of semantic bias is estimated by weighting augmented samples in the embedding space. Specifically, the samples are mixed in both sample space and embedding space respectively, to construct more reliable and task-relevant positive sample pairs. Then, a patch-mixing contrastive objective is designed to indicate the magnitude of semantic bias by utilizing a mixed embedding weighted by virtual soft labels. Extensive experiments were conducted, demonstrating that <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$p$</tex-math></inline-formula> -Mix significantly outperforms current state-of-the-art approaches. Especially, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$p$</tex-math></inline-formula> -Mix achieves an accuracy of 91.95% on the CIFAR-10 benchmark with only 2 labels available for each category, which exceeds the second-best method ICL-SSL by 3.22%. Source code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/XiaochangHu/P-Mix</uri>

A novel full-convolution UNet-transformer for medical image segmentation

The Transformer-based methods are still unable to effectively model local contexts although they make up for the deficiency of remote information dependencies for approaches based on small kernel CNNs. To overcome such a shortage, this paper proposes a novel full-convolution UNet Transformer model, FC-UNETTR, for medical image segmentation. First, a novel global–local attention module is proposed by utilizing multiple small kernels of different sizes for depth-wise convolutions to expand the receptive field of the network model, increase the remote dependence of semantic information in the encoder stage, and also improve the feature extraction capability of the network for fuzzy edges. Then, a reparametrized feedforward network is developed to further improve the local information extraction and mitigate the coupling between feature maps such that the relationship between feature map channels can be better revealed. Furthermore, the skip connection and decoder are redesigned by constructing a dense multiscale module instead of traditional ResNet modules to mitigate semantic bias. Benefiting from the above improvements, the constructed FC-UNETTR without pre-training demonstrates a strong capability to extract local features and capture long-range dependencies of images in medical image segmentation. Experiments show that FC-UNETTR achieves an excellent performance of 85.67% for DSC and 7.82 for HD metrics on the Synapse dataset with fewer model parameters compared with state-of-the-art networks. Furthermore, DSC reaches 92.46% and 94.76% on the ACDC dataset and the private dataset of oral graft bone, respectively, outperforming some of the latest medical image segmentation models..

Similarity Graph-correlation Reconstruction Network for unsupervised cross-modal hashing

Existing cross-modal hash retrieval methods can simultaneously enhance retrieval speed and reduce storage space. However, these methods face a major challenge in determining the similarity metric between two modalities. Specifically, the accuracy of intra-modal and inter-modal similarity measurements is inadequate, and the large gap between modalities leads to semantic bias. In this paper, we propose a Similarity Graph-correlation Reconstruction Network (SGRN) for unsupervised cross-modal hashing. Particularly, the local relation graph rebasing module is used to filter out graph nodes with weak similarity and associate graph nodes with strong similarity, resulting in fine-grained intra-modal similarity relation graphs. The global relation graph reconstruction module is further strengthens cross-modal correlation and implements fine-grained similarity alignment between modalities. In addition, in order to bridge the modal gap, we combine the similarity representation of real-valued and hash features to design the intra-modal and inter-modal training strategies. SGRN conducted extensive experiments on two cross-modal retrieval datasets, and the experimental results effectively validated the superiority of the proposed method and significantly improved the retrieval performance.

Twin Pseudo-training for semi-supervised semantic segmentation

Recent advancements in semi-supervised semantic segmentation have demonstrated the effectiveness of utilizing pseudo labels supervision to mitigate the limitations of pixel-wise annotations. However, pseudo-labels generated using self-training techniques typically contain a significant amount of noise, which can impede the training process of the supervised model. In this study, we identify low- and high-level semantic errors as the two key factors that hinder the accuracy of pseudo labels. To fully exploit the potential of pseudo labels, we introduce a novel semi-supervised framework named Twin Pseudo-training (TPseudo), which employs a consistency and disagreement collaboration strategy. Specifically, we correct pseudo labels with a False-positive Filter (FPF) to reduce high-level semantic noise and refine low-level semantic biases using a Semantic Error Detector (SED). Lastly, we design a Self-Adaptive Weight (SAW) loss function based on a disagreement between two predictions to exploit each pixel of pseudo labels. Experimental results on the standard benchmarks PASCAL VOC2012 and Cityscapes demonstrate the efficacy of the proposed method.

Camouflaged object detection with counterfactual intervention

Camouflaged object detection (COD) aims to identify camouflaged objects hiding in their surroundings, which is a valuable yet challenging task. The main challenge is that there are ambiguous semantic biases in the camouflaged object datasets, which affect the results of COD. To address this challenge, we design a counterfactual intervention network (CINet) to mitigate the influences of ambiguous semantic biases and obtain accurate COD. Specifically, our CINet consists of three key modules, i.e., texture-aware interaction module (TIM), context-aware fusion module (CFM), and counterfactual intervention module (CIM). The TIM is designed to extract the refined textures for accurate localization, the CFM is proposed to fuse the multi-scale contextual features to enhance the detection performance, and the CIM is presented to learn more effective textures and make unbiased predictions. Unlike most existing COD methods that directly capture contextual features through the final loss function, we develop a counterfactual intervention strategy to learn more effective contextual textures. Extensive experiments on four challenging benchmark datasets demonstrate that our CINet significantly outperforms 31 state-of-the-art methods.

Cognitive Network Science Reveals Bias in GPT-3, GPT-3.5 Turbo, and GPT-4 Mirroring Math Anxiety in High-School Students

Large Language Models (LLMs) are becoming increasingly integrated into our lives. Hence, it is important to understand the biases present in their outputs in order to avoid perpetuating harmful stereotypes, which originate in our own flawed ways of thinking. This challenge requires developing new benchmarks and methods for quantifying affective and semantic bias, keeping in mind that LLMs act as psycho-social mirrors that reflect the views and tendencies that are prevalent in society. One such tendency that has harmful negative effects is the global phenomenon of anxiety toward math and STEM subjects. In this study, we introduce a novel application of network science and cognitive psychology to understand biases towards math and STEM fields in LLMs from ChatGPT, such as GPT-3, GPT-3.5, and GPT-4. Specifically, we use behavioral forma mentis networks (BFMNs) to understand how these LLMs frame math and STEM disciplines in relation to other concepts. We use data obtained by probing the three LLMs in a language generation task that has previously been applied to humans. Our findings indicate that LLMs have negative perceptions of math and STEM fields, associating math with negative concepts in 6 cases out of 10. We observe significant differences across OpenAI’s models: newer versions (i.e., GPT-4) produce 5× semantically richer, more emotionally polarized perceptions with fewer negative associations compared to older versions and N=159 high-school students. These findings suggest that advances in the architecture of LLMs may lead to increasingly less biased models that could even perhaps someday aid in reducing harmful stereotypes in society rather than perpetuating them.

TDRLM: Stylometric learning for authorship verification by Topic-Debiasing

Authorship verification is the task of determining if a candidate is the real author of the provided sequences of texts. Stylometric learning is the study of the writing style by analyzing the textual data through computational approaches for many linguistic tasks such as authorship verification, authorship identification, etc. In recent years, Deep learning has been used as an effective solution to stylometric learning for authorship verification. However, most of the existing studies do not consider the semantic biases captured by the model, and the learned stylometric features include a strong indicator of the topical biases. During the extraction of stylometric features, we found that specific words related to the latent topics impact the performance of machine learning algorithms and, therefore, are not generalizable toward out-of-sample authors. In this paper, we propose a latent topic score dictionary with a Topic-Debiasing Representation Learning Model (TDRLM) for stylometric representation learning on the problem of authorship verification. The model applies position-specific topic scores on a topic-debiasing attention mechanism in order to adjust the tokenized texts based on the topical bias. The experimental results show that our proposed approach achieves the best performance with the highest Area Under Curve (AUC) of 92.47% for the Twitter-Foursquare dataset and 93.11% for the (ICWSM) Twitter dataset, both better than the state-of-the-art stylometric learning models and the up-trained language models.

TCU-Net: Transformer Embedded in Convolutional U-Shaped Network for Retinal Vessel Segmentation.

Optical coherence tomography angiography (OCTA) provides a detailed visualization of the vascular system to aid in the detection and diagnosis of ophthalmic disease. However, accurately extracting microvascular details from OCTA images remains a challenging task due to the limitations of pure convolutional networks. We propose a novel end-to-end transformer-based network architecture called TCU-Net for OCTA retinal vessel segmentation tasks. To address the loss of vascular features of convolutional operations, an efficient cross-fusion transformer module is introduced to replace the original skip connection of U-Net. The transformer module interacts with the encoder's multiscale vascular features to enrich vascular information and achieve linear computational complexity. Additionally, we design an efficient channel-wise cross attention module to fuse the multiscale features and fine-grained details from the decoding stages, resolving the semantic bias between them and enhancing effective vascular information. This model has been evaluated on the dedicated Retinal OCTA Segmentation (ROSE) dataset. The accuracy values of TCU-Net tested on the ROSE-1 dataset with SVC, DVC, and SVC+DVC are 0.9230, 0.9912, and 0.9042, respectively, and the corresponding AUC values are 0.9512, 0.9823, and 0.9170. For the ROSE-2 dataset, the accuracy and AUC are 0.9454 and 0.8623, respectively. The experiments demonstrate that TCU-Net outperforms state-of-the-art approaches regarding vessel segmentation performance and robustness.

To make yourself invisible with Adversarial Semantic Contours

Modern object detectors are vulnerable to adversarial examples, which may bring risks to real-world applications. The sparse attack is an important task which, compared with the popular adversarial perturbation on the whole image, needs to select the potential pixels that is generally regularized by an $\ell_0$-norm constraint, and simultaneously optimize the corresponding texture. The non-differentiability of $\ell_0$ norm brings challenges and many works on attacking object detection adopted manually-designed patterns to address them, which are meaningless and independent of objects, and therefore lead to relatively poor attack performance. In this paper, we propose Adversarial Semantic Contour (ASC), an MAP estimate of a Bayesian formulation of sparse attack with a deceived prior of object contour. The object contour prior effectively reduces the search space of pixel selection and improves the attack by introducing more semantic bias. Extensive experiments demonstrate that ASC can corrupt the prediction of 9 modern detectors with different architectures (\e.g., one-stage, two-stage and Transformer) by modifying fewer than 5\% of the pixels of the object area in COCO in white-box scenario and around 10\% of those in black-box scenario. We further extend the attack to datasets for autonomous driving systems to verify the effectiveness. We conclude with cautions about contour being the common weakness of object detectors with various architecture and the care needed in applying them in safety-sensitive scenarios.