Attentional Bottleneck Research Articles

MTANet: Multi-Task Attention Network for Automatic Medical Image Segmentation and Classification.

Medical image segmentation and classification are two of the most key steps in computer-aided clinical diagnosis. The region of interest were usually segmented in a proper manner to extract useful features for further disease classification. However, these methods are computationally complex and time-consuming. In this paper, we proposed a one-stage multi-task attention network (MTANet) which efficiently classifies objects in an image while generating a high-quality segmentation mask for each medical object. A reverse addition attention module was designed in the segmentation task to fusion areas in global map and boundary cues in high-resolution features, and an attention bottleneck module was used in the classification task for image feature and clinical feature fusion. We evaluated the performance of MTANet with CNN-based and transformer-based architectures across three imaging modalities for different tasks: CVC-ClinicDB dataset for polyp segmentation, ISIC-2018 dataset for skin lesion segmentation, and our private ultrasound dataset for liver tumor segmentation and classification. Our proposed model outperformed state-of-the-art models on all three datasets and was superior to all 25 radiologists for liver tumor diagnosis.

Lightweight Sheep Head Detection and Dynamic Counting Method Based on Neural Network.

To achieve rapid and precise target counting, the quality of target detection serves as a pivotal factor. This study introduces the Sheep's Head-Single Shot MultiBox Detector (SH-SSD) as a solution. Within the network's backbone, the Triple Attention mechanism is incorporated to enhance the MobileNetV3 backbone, resulting in a significant reduction in network parameters and an improvement in detection speed. The network's neck is constructed using a combination of the Spatial Pyramid Pooling module and the Triple Attention Bottleneck module. This combination enhances the extraction of semantic information and the preservation of detailed feature map information, with a slight increase in network parameters. The network's head is established through the Decoupled Head module, optimizing the network's prediction capabilities. Experimental findings demonstrate that the SH-SSD model attains an impressive average detection accuracy of 96.11%, effectively detecting sheep's heads within the sample. Notably, SH-SSD exhibits enhancements across various detection metrics, accompanied by a significant reduction in model parameters. Furthermore, when combined with the DeepSort tracking algorithm, it achieves high-precision quantitative statistics. The SH-SSD model, introduced in this paper, showcases commendable performance in sheep's head detection and offers deployment simplicity, thereby furnishing essential technical support for the advancement of intelligent animal husbandry practices.

Multi-Modal Representation via Contrastive Learning with Attention Bottleneck Fusion and Attentive Statistics Features.

The integration of information from multiple modalities is a highly active area of research. Previous techniques have predominantly focused on fusing shallow features or high-level representations generated by deep unimodal networks, which only capture a subset of the hierarchical relationships across modalities. However, previous methods are often limited to exploiting the fine-grained statistical features inherent in multimodal data. This paper proposes an approach that densely integrates representations by computing image features' means and standard deviations. The global statistics of features afford a holistic perspective, capturing the overarching distribution and trends inherent in the data, thereby facilitating enhanced comprehension and characterization of multimodal data. We also leverage a Transformer-based fusion encoder to effectively capture global variations in multimodal features. To further enhance the learning process, we incorporate a contrastive loss function that encourages the discovery of shared information across different modalities. To validate the effectiveness of our approach, we conduct experiments on three widely used multimodal sentiment analysis datasets. The results demonstrate the efficacy of our proposed method, achieving significant performance improvements compared to existing approaches.

LKDA-GAN: Cross-modality image synthesis via Generative Adversarial Network aggregating large kernel decomposable attention bottleneck block

Background and Objective:Mainstream image synthesis methods fail to capture local contextual information long-range dependence, and adaptability, especially under the influence of computation overload and random noise distribution, leading to the loss of contrast and detailed information. Approach:To alleviate these issues, we propose a Generative Adversarial Network aggregating large kernel decomposable attention (LKDA-GAN) bottleneck block for cross-modality image synthesis. Initially, a novel LKDA module is proposed by combining a spatial local convolution, a spatial long-range convolution, and a channel convolution, which aims to balance the local contextual information and the long-range dependence, and enhance the channel adaptability by enlarging the receptive field. Subsequently, a bottleneck block is designed and integrated into the LKDA module by feature dimensional transformation to capture rich semantic information and alleviate computational overhead. Ultimately, an auxiliary registration network (ARN) based on the noise transition matrix is put forward to learn the prior knowledge of the noise distribution and produce the unique optimal solution. Furthermore, it is based on Res-UNet to avoid network degradation. Main results:By analyzing qualitative assessment and quantitative measurement, extensive experiments demonstrate the proposed approach outperforms the state-of-the-art methods, and ablation studies also show the superiority of LKDA. Significance:LKDA-GAN provides a suitable way to synthesize images between different modalities, which is conductive to indicate disease areas and improve diagnosis accuracy.

TransBoNet: Learning camera localization with Transformer Bottleneck and Attention

6DoF camera localization is an important component of autonomous driving and navigation. Deep learning has achieved impressive results in localization, but its robustness in dynamic environments has not been adequately addressed. In this paper, we propose a framework based on hybrid attention mechanism which can be generally applied to existing CNN-based pose regressors to improve their robustness in dynamic environments. Specifically, we propose a novel Transformer Bottleneck (TBo) block including convolution, channel attention, and a position-aware self-attention mechanism, which extracts more geometrically robust features by capturing the corresponding long-term dependencies between pixels. Furthermore, we introduce shuffle attention (SA) before the pose regressor, which integrates feature information in both spatial and channel dimensions, forcing the network to learn geometrically robust features, reducing the effects of dynamic objects and illumination conditions to improve camera localization accuracy. We evaluate our method on commonly benchmarked indoor and outdoor datasets and the experimental results show that our proposed method can significantly improve localization performance compared compare favorably to contemporary pose regressors schemes. In addition, extensive ablation evaluations are conducted to prove the effectiveness of our proposed hybrid attention bottleneck block for pose regression networks.

Pixel-Wise Grasp Detection via Twin Deconvolution and Multi-Dimensional Attention

The grasp detection is crucial to high-quality robotic grasping. Typically, the mainstream encoder-decoder regression solution is attractive due to its high accuracy and efficiency, however, it is still challenging to solve the checkerboard artifacts from the uneven overlap of convolution results in decoder, and features from the encoder also need to be further refined. In this paper, a novel pixel-wise grasp detection network is proposed, which is composed of an encoder, a multi-dimensional attention bottleneck, and a decoder based on twin deconvolution. The proposed decoder introduces a twin branch upon the original transposed convolution branch. Through the overlap degree matrix provided by the twin branch, the original branch is re-weighted and then the checkerboard artifacts of the original branch are eliminated. Besides, to deeply explore the intrinsic relationship of features and strengthen feature discrimination, residual multi-head self-attention, cross-amplitude attention, and channel attention are integrated together. As a result, adaptive feature refinement is achieved. The effectiveness of the proposed method is verified by experiments.

Serial attention to serial memory: The psychological refractory period in forward and backward cued recall

Guided by the conjecture that memory retrieval is attention turned inward, we examined serial attention in serial memory, combining the psychological refractory period (PRP) procedure from attention research with cued recall of two items from brief six-item lists. We report six experiments showing robust PRP effects in cued recall from memory (1–4) and cued report from perceptual displays (5–6), which suggest that memory retrieval requires the same attentional bottleneck as “retrieval” from perception. There were strong direction effects in each memory experiment. Response time (RT) was shorter and accuracy was higher when the cues occurred in the forward direction (left-to-right, top-to-bottom, first-to-last), replicating differences between forward and backward serial recall. Cue positions had strong effects on RT and accuracy in the memory experiments (1–4). The pattern suggested that subjects find cued items in memory by stepping through the list from the beginning or the end, with a preference for starting at the beginning. The perceptual experiments (5–6) showed weak effects of position that were more consistent with direct access. In all experiments, the distance between the cues in the list (lag) had weak effects, suggesting that subjects searched for each cue from the beginning or end of the list more often than they moved through the list from the first cue to the second. Direction, distance, and lag effects on RT and inter-response interval changed with SOA in a manner that suggested they affect bottleneck or pre-bottleneck processes that create and execute a plan for successive retrievals. We conclude that sequential retrieval from memory and sequential attention to perception engage the same computations and we show how computational models of memory can be interpreted as models of attention focused on memory.

MFVC: Urban Traffic Scene Video Caption Based on Multimodal Fusion

With the development of electronic technology, intelligent cars can gradually realize more complex artificial intelligence algorithms. The video caption algorithm is one of them. However, current video caption algorithms only consider single-visual information when applied to urban traffic scenes, which leads to the failure to generate accurate captions of complex sets. The multimodal fusion algorithm based on Transformer is one of the solutions to this problem. However, the existing algorithms have the difficulties of a low fusion performance and high computational complexity. We propose a new video caption Transformer-based model, the MFVC (Multimodal Fusion for Video Caption), to solve these issues. We introduce audio modal data and the attention bottleneck module to increase the available information to describe the generative model and improve the model effect with less operation costs through the attention bottleneck module. Finally, the experiment is conducted on the available datasets, MSR-VTT and MSVD. Meanwhile, to verify the effect of the model on the urban traffic scene, the experiment is carried out on the self-built traffic caption dataset BUUISE, and the evaluation index confirms the model. This model can achieve good results on both available datasets and urban traffic datasets and has excellent application prospects in the intelligent driving industry.

E-GCS: Detection of COVID-19 through classification by attention bottleneck residual network

Background:Recently, the coronavirus disease 2019 (COVID-19) has caused mortality of many people globally. Thus, there existed a need to detect this disease to prevent its further spread. Hence, the study aims to predict COVID-19 infected patients based on deep learning (DL) and image processing. Objectives:The study intends to classify the normal and abnormal cases of COVID-19 by considering three different medical imaging modalities namely ultrasound imaging, X-ray images and CT scan images through introduced attention bottleneck residual network (AB-ResNet). It also aims to segment the abnormal infected area from normal images for localising the disease infected area through the proposed edge based graph cut segmentation (E-GCS). Methodology:AB-ResNet is used for classifying images whereas E-GCS segment the abnormal images. The study possess various advantages as it rely on DL and possess capability for accelerating the training speed of deep networks. It also enhance the network depth leading to minimum parameters, minimising the impact of vanishing gradient issue and attaining effective network performance with respect to better accuracy. Results/Conclusion:Performance and comparative analysis is undertaken to evaluate the efficiency of the introduced system and results explores the efficiency of the proposed system in COVID-19 detection with high accuracy (99%).

Attention Oligopoly

We model digital platforms as attention brokers that have proprietary information about their users’ product preferences and sell targeted ad space to retail product industries. Retail producers—incumbents or entrants—compete for access to this attention bottleneck. We discuss when increased concentration among attention brokers results in a tightening of the attention bottleneck, leading to higher ad prices, fewer ads being sold to entrants, and lower consumer welfare in the product industries. The welfare effect is characterized in terms of patterns of individual usage across platforms. A merger assessment that relies on aggregate platform usage alone can be highly biased. (JEL D43, D83, G34, L81, L86, M37)

Inter-Individual Differences in Executive Functions Predict Multitasking Performance - Implications for the Central Attentional Bottleneck.

Human multitasking suffers from a central attentional bottleneck preventing parallel performance of central mental operations, leading to profound deferments in task performance. While previous research assumed that the deferment is caused by a mere waiting time (refractory period), we show that the bottleneck requires executive functions (EF; active scheduling account) accounting for a profound part of the deferment. Three participant groups with EF impairments (dyslexics, highly neurotics, deprived smokers) showed worse multitasking performance than respective control groups. Three further groups with EF improvements (video-gamers, bilinguals, coffee consumers) showed improved multitasking. Finally, three groups performed a dual-task and different measures of EF (reading span, rotation span, symmetry span) and showed significant correlations between multitasking performance and working memory capacity. Demands on EF during multitasking may cause more errors, mental fatigue and stress, with parts of the population being considerably more prone to this.

The importance of an exaggerated attention bottleneck for understanding psychopathy.

The psychopath has long captured the imagination. A name such as Ted Bundy evokes a morbid curiosity. The crimes committed by Bundy are so cruel that it is hard to imagine how someone could do such things. In this review we discuss evidence that exaggeration in an attention bottleneck is one mechanism that makes it possible for psychopathic individuals to be adept at focusing on a single stimulus feature or goal but struggle to process multiple streams of information simultaneously. This exaggeration may partly explain the behavioral, affective, and social deficits that are apparent among psychopathic individuals. Further research on this attentional mechanism may promote a science that adequately captures the complexity of psychopathic behavior and offers new avenues for intervention.

Capacity-free automatic processing of facial expressions of emotion.

Are facial expressions of emotion processed automatically? Some authors have not found this to be the case (Tomasik et al., 2009). Here we revisit the question with a novel experimental logic-the backward correspondence effect (BCE). In three dual-task studies, participants first categorized a sound (Task 1) and then indicated the location of a target face (Task 2). In Experiment 1, Task 2 required participants to search for one facial expression of emotion (angry or happy). We observed positive BCEs, indicating that facial expressions of emotion bypassed the central attentional bottleneck and thus were processed in a capacity-free, automatic manner. In Experiment 2, we replicated this effect but found that morphed emotional expressions (which were used by Tomasik) were not processed automatically. In Experiment 3, we observed similar BCEs for another type of face processing previously shown to be capacity-free: identification of familiar faces (Jung et al., 2013). We conclude that facial expressions of emotion are identified automatically when sufficiently unambiguous. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

The working memory costs of a central attentional bottleneck in multitasking

When two (or more) tasks, each requiring a rapid response, are performed at the same time then serial processing may occur at certain processing stages, such as the response selection. There is accumulating evidence that such serial processing involves additional control processes, such as inhibition, switching, and scheduling (termed the active scheduling account). The present study tested whether the existence of serial processing in multitasking leads to a requirement for processes that coordinate processing in this way (active scheduling account) and, furthermore, whether such control processes are linked to the executive functions (EF) of working memory (WM). To test this question, we merged the psychological refractory period (PRP) paradigm with a WM task, creating a complex WM span task. Participants were presented with a sequence of letters to remember, followed by a processing block in which they had to perform either a single task or a dual task, and finally were asked to recall the letters. Results showed that WM performance, i.e. the amount of letters recalled in the correct order, decreased when performing a dual task as compared to performing a single task during the retention interval. Two further experiments supported this finding using manipulations of the dual task difficulty. We conclude that the existence of serial processing in multitasking demands additional control processes (active scheduling) and that these processes are strongly linked to the executive functions of working memory.

Attentive and context-aware deep network for saliency prediction on omni-directional images

Understanding visual attention of observers on omni-directional images gains interest along with the booming trend of virtual reality applications. In this paper, we propose a novel attentive and context-aware network for saliency prediction on omni-directional images, which is named as ACSalNet. In this architecture, considering the problem of insufficient receptive fields of high-level features, a Deformable Attention Bottleneck (DAB) is first proposed to strengthen the high-level feature extractor and effectively focus the limited receptive field of the model to the key areas. Then, to reduce the semantic gap between features of different levels and introduce context-aware information, we further design a Context-aware Feature Pyramid Module (CFPM). In the testing phase, in order to reduce the error of prediction directly on the equirectangular images while retaining their integrity, a novel projection method called Multiple Sphere Rotation (MSR) is proposed. Extensive experiments illustrate that the proposed method outperforms the state-of-the-art models under different evaluation metrics on the public saliency benchmarks.

Towards an Interpretable Deep Driving Network by Attentional Bottleneck

Deep neural networks are a key component of behavior prediction and motion generation for self-driving cars. One of their main drawbacks is a lack of transparency: they should provide easy to interpret rationales for what triggers certain behaviors. We propose an architecture called Attentional Bottleneck with the goal of improving transparency. Our key idea is to combine visual attention, which identifies what aspects of the input the model is using, with an information bottleneck that enables the model to only use aspects of the input which are important. This not only provides sparse and interpretable attention maps (e.g. focusing only on specific vehicles in the scene), but it adds this transparency at no cost to model accuracy. In fact, we find improvements in accuracy when applying Attentional Bottleneck to the ChauffeurNet model, whereas we find that the accuracy deteriorates with a traditional visual attention model.

Preserved sensory processing but hampered conflict detection when stimulus input is task-irrelevant.

Conflict detection in sensory input is central to adaptive human behavior. Perhaps unsurprisingly, past research has shown that conflict may even be detected in the absence of conflict awareness, suggesting that conflict detection is an automatic process that does not require attention. To test the possibility of conflict processing in the absence of attention, we manipulated task relevance and response overlap of potentially conflicting stimulus features across six behavioral tasks. Multivariate analyses on human electroencephalographic data revealed neural signatures of conflict only when at least one feature of a conflicting stimulus was attended, regardless of whether that feature was part of the conflict, or overlaps with the response. In contrast, neural signatures of basic sensory processes were present even when a stimulus was completely unattended. These data reveal an attentional bottleneck at the level of objects, suggesting that object-based attention is a prerequisite for cognitive control operations involved in conflict detection.

Psychopathy is associated with an exaggerated attention bottleneck: EEG and behavioral evidence from a dual-task paradigm

Psychopathy is a personality disorder associated with a chronic disregard for the welfare of others. The attention bottleneck model of psychopathy asserts that the behavior of individuals higher on psychopathy is due to an exaggerated attention bottleneck that constrains all information processing, regardless of the information's potential goal-relevance. To date, the majority of research on the attention bottleneck model of psychopathy conceptually applied the tenets of the model but did not implement methods that directly test an exaggeration of the bottleneck in psychopathy. Accordingly, the presence of an exaggerated bottleneck, the exact expression of that bottleneck, and its potential mechanistic relevance for behavior in individuals higher on psychopathy remains untested. To address these gaps, a sample of 78 male community members, evaluated for psychopathic traits using the Self-Report Psychopathy-III scale, completed an EEG-based dual-task paradigm examining short stimulus onset asynchrony (SOA; 300 ms), long SOA (1,100 ms), and single-task baseline conditions. Additionally, participants were asked about their frequency of real-world risky, impulsive, and antisocial behaviors. Psychopathy was associated with slower reaction times to second targets (T2s) presented during the dual-task conditions, relative to the baseline condition. Psychopathy also was associated with blunted P300 responses, a neural index of stimulus evaluation, across all types of T2 events. Finally, bottleneck-related interference during the short SOA events mediated the relationship between psychopathy and real-world behavior. These findings suggested that individuals higher on psychopathy exhibit an exaggerated bottleneck which produces intense and long-lasting interference, impacting all information processing and partially contributing to their maladaptive behavior.

OCPAN: multi‐person pose estimation with more attention on residual information

Redundant information generated in the process of feature fusion and residual information always remain challenging in multi-person pose estimation. In this Letter, the authors present an optimised cascaded pyramid attention network composed of two novel modules to reduce the redundant information and highlight the residual information for more accurate results. The first module, called channel optimisation module (COM), optimises the channel at different levels. The proposed COM consisting of convolution layers and two pooling layers performs on the feature maps Conv-2 ∼ 5 of different resolution to reduce the redundant information after convolution operation. The second module, called residual attention bottleneck module (RABM), highlights the residual information in the bottleneck. Through adopting the attention mechanism in the bottleneck, RABM performs a significant role in reducing the redundant information. In contrast to the previous work, they attach more attention to residual information, helping to tackle the challenging problem for the ‘hard’ keypoints. By merging COM and RABM into Cascaded Pyramid Network, the authors' method achieves competitive results of 74.6 AP on the COCO keypoint benchmark, outperforming the state-of-the-art results.

Ideomotor compatibility enables automatic response selection.

A task is ideomotor (IM)-compatible when there is high conceptual similarity between the stimulus and the associated response (e.g., pressing a left key when an arrow points to the left). For such an easy task, can response selection operate automatically, bypassing the attentional bottleneck that normally constrains dual-task performance? To address this question, we manipulated the IM compatibility of a Task 2 that was performed concurrently with a non-IM-compatible Task 1, using the psychological refractory period procedure. Single-task trials, randomly intermixed with dual-task trials, served as a baseline against which to assess dual-task costs. The results indicated bottleneck bypassing (i.e., simultaneous response selection on both tasks) when Task 2 was IM-compatible, as evidenced by negligible dual-task costs on Task 2 (as well as on Task 1), very high percentages of response reversals, and weak correlations between Task-1 and Task-2 reaction times. These findings were supported by a fine-grained simulation analysis of inter-response intervals. We conclude that the perception of an IM-compatible stimulus directly activates the response code, which can then be selecting automatically, without recruiting central attention, consistent with A. G. Greenwald's (Journal of Experimental Psychology, 94, 52-57, 1972) original theory of IM compatibility.