Search

Contrastive learning of sentence representations has achieved great improvements in several natural language processing tasks. However, the supervised contrastive learning model trained on the natural language inference (NLI) dataset is insufficient to elucidate the semantics of sentences since it is prone to make a prediction based on heuristics. Herein, by using the ParsEVAL and the word overlap metric, it is shown that sentence pairs in the NLI dataset have strong syntactic similarity and propose a framework to compensate for this problem in two aspects. 1) Apply simple syntactic transformations to the hypothesis and 2) expand the objective to SupCon Loss to leverage variants of sentences. The method is evaluated on semantic textual similarity (STS) tasks and transfer tasks. The proposed methods improve the performance of the BERT‐based baseline in STS Benchmark and SICK Relatedness by 1.48% and 2.2%. Furthermore, the model achieves 82.65% on the HANS benchmark dataset, to the best of our knowledge, which is a state‐of‐the‐art performance demonstrating that our approach is effective in grasping semantics without heuristics in the NLI dataset at supervised contrastive learning. The code is available at https://github.com/whnhch/Break-the-Similarity.

Recent studies on image super-resolution (SR) have focused on just expanding the receptive field. However, the insight from local attribution maps has enlightened the need for an improved exploitation of information within the receptive field. To address this, we propose a novel image super-resolution model, named Uniwin, that balances local and global interactions through unifying two types of window-based local attention mechanisms: shifted-window attention and sliding-window attention. Uniwin combines swift global context access with comprehensive local context capture. Our approach involves the initial sliding-window attention to collect comprehensive local pattern information, followed by the non-overlapping shifted-window attention to expand the receptive field for global interactions. Empirical evaluations demonstrate Uniwin’s superiority over state-of-the-art models across five benchmark datasets. Specifically, in SR $\times 2$ task, our model achieved 0.25dB higher PSNR on the Set14 dataset, and 0.14dB higher PSNR on the Urban100 dataset than existing state-of-the-art model using a similar number of parameters. Additionally, our model achieved comparable performance to existing large-scale models with only 58% of the parameters. Ablation studies on sliding-window and shifted-window attention mechanisms reveal the critical importance of the context harmonization for image super-resolution. In conclusion, Uniwin represents a reasonable solution that effectively integrates global and local attention mechanisms, enhancing image super-resolution performance.

Anomaly detection has gained significant attention with the advancements in deep neural networks. Effective training requires both normal and anomalous data, but this often leads to a class imbalance, as anomalous data is scarce. Traditional augmentation methods struggle to maintain the correlation between anomalous patterns and their surroundings. To address this, we propose an adjacent augmentation technique that generates synthetic anomaly images, preserving object shapes while distorting contours to enhance correlation. Experimental results show that adjacent augmentation captures high-quality anomaly features, achieving superior AU-ROC and AU-PR scores compared to existing methods. Additionally, our technique produces synthetic normal images, aiding in learning detailed normal data features and reducing sensitivity to minor variations. Our framework considers all training images within a batch as positive pairs, pairing them with synthetic normal images as positive pairs and with synthetic anomaly images as negative pairs. This compensates for the lack of anomalous features and effectively distinguishes between normal and anomalous features, mitigating class imbalance. Using the ResNet50 network, our model achieved perfect AU-ROC and AU-PR scores of 100% in the bottle category of the MVTec-AD dataset. We are also investigating the relationship between anomalous pattern size and detection performance.

Background The aging population in Korea has driven a surge in demand for elderly care services, leading to significant growth in elderly welfare facilities, particularly Adult Daycare Centers (ADCs). However, despite advancements in care facilities, caregivers continue to face challenges in providing suitable elderly care due to difficulties arising from gaps in the latest information on the elderly and their coping abilities. Objective The objective of this study is to develop and evaluate the effectiveness of the elderly care assistant system, which facilitates the sharing of information and knowledge necessary for elderly care among caregivers. Methods The ECA system was designed to support knowledge sharing through a knowledge management system based on an ontological knowledge model, with a web-based user interface for improved accessibility. A field trial was conducted at ADC in Seoul from August 17 to September 21, with eight caregivers participating. A mixed-methods approach, involving both surveys and interviews, was employed to gauge the ECA system's effectiveness. Results The study found that the use of the ECA was beneficial in promoting knowledge sharing among caregivers. Additionally, caregivers noted the potential benefits of using the ECA in conjunction with family caregivers, who can offer additional information and perspectives on elderly care. Conclusions This study presents preliminary evidence of the potential benefits of a care knowledge sharing system among various caregivers in elderly care. Although the elderly care assistant effectively promotes knowledge sharing, more research is needed to fully understand its impact on elderly care outcomes.

As the performance of computing devices such as graphics processing units (GPUs) has improved dramatically, many deep neural network models, especially convolutional neural networks (CNNs), have been widely applied to various applications such as image classification, semantic segmentation, and object recognition. However, effective first-order optimization methods for CNNs have rarely been studied, although many CNN models have been successfully developed. Accordingly, this paper investigates various advanced adaptive solution search methods and proposes a new first-order optimization algorithm for CNNs called Adam-ASC. Our approach uses four sophisticated adaptive solution search methods to adjust its search strength in the complicated large-dimensional weight solution space spanned by a loss function. At the same time, we explain how they can be combined compensatively to form a complete optimizer with a detailed implementation. From the experiments, we found that our Adam-ASC can significantly improve the image recognition performance of practical CNNs in both the image classification and segmentation tasks. These experimental results show that the four fundamental methods of Adam-ASC and their compensative combination strategy play a crucial role in training CNNs by effectively finding their optimal weights.

Transformer-based networks using attention mechanisms have shown promising results in low-level vision tasks, such as image super-resolution (SR). Specifically, recent studies that utilize window-based self-attention mechanisms have exhibited notable advancements in image SR. However, window-based self-attention, results in a slower expansion of the receptive field, thereby restricting the modeling of long-range dependencies. To address this issue, we introduce a novel dilated window transformer, namely DWT, which utilizes a dilation strategy. We employ a simple yet efficient dilation strategy that enlarges the window by inserting intervals between the tokens of each window to enable rapid and effective expansion of the receptive field. In particular, we adjust the interval between the tokens to become wider as the layers go deeper. This strategy enables the extraction of local features by allowing interaction between neighboring tokens in the shallow layers while also facilitating efficient extraction of global features by enabling interaction between not only adjacent tokens but also distant tokens in the deep layers. We conduct extensive experiments on five benchmark datasets to demonstrate the superior performance of our proposed method. Our DWT surpasses the state-of-the-art network of similar sizes by a PSNR margin of 0.11dB to 0.27dB on the Urban100 dataset. Moreover, even when compared to state-of-the-art network with about 1.4 times more parameters, DWT achieves competitive results for both quantitative and visual comparisons.

Most thickness measurement techniques using X-ray radiation are unsuitable in field processes involving fast-moving organic films. Herein, we propose a Compton scattering X-ray radiation method, which probes the light elements in organic materials, and a new simple, non-destructive, and non-contact calibration-free real-time film thickness measurement technique by setting up a bench-top X-ray thickness measurement system simulating a field process dealing with thin flexible organic films. The use of X-ray fluorescence and Compton scattering X-ray radiation reflectance signals from films in close contact with a roller produced accurate thickness measurements. In a high-thickness range, the contribution of X-ray fluorescence is negligible, whereas that of Compton scattering is negligible in a low-thickness range. X-ray fluorescence and Compton scattering show good correlations with the organic film thickness (R2 = 0.997 and 0.999 for X-ray fluorescence and Compton scattering, respectively, in the thickness range 0–0.5 mm). Although the sensitivity of X-ray fluorescence is approximately 4.6 times higher than that of Compton scattering, Compton scattering signals are useful for thick films (e.g., thicker than ca. 1–5 mm under our present experiment conditions). Thus, successful calibration-free thickness monitoring is possible for fast-moving films, as demonstrated in our experiments.

Abstract With recent advances in biotechnology and sequencing technology, the microbial community has been intensively studied and discovered to be associated with many chronic as well as acute diseases. Even though a tremendous number of studies describing the association between microbes and diseases have been published, text mining methods that focus on such associations have been rarely studied. We propose a framework that combines machine learning and natural language processing methods to analyze the association between microbes and diseases. A hierarchical long short-term memory network was used to detect sentences that describe the association. For the sentences determined, two different parse tree-based search methods were combined to find the relation-describing word. The ensemble model of constituency parsing for structural pattern matching and dependency-based relation extraction improved the prediction accuracy. By combining deep learning and parse tree-based extractions, our proposed framework could extract the microbe-disease association with higher accuracy. The evaluation results showed that our system achieved an F-score of 0.8764 and 0.8524 in binary decisions and extracting relation words, respectively. As a case study, we performed a large-scale analysis of the association between microbes and diseases. Additionally, a set of common microbes shared by multiple diseases were also identified in this study. This study could provide valuable information for the major microbes that were studied for a specific disease. The code and data are available at https://github.com/DMnBI/mdi_predictor .

The aim of this study was to present an optimal diagnostic protocol by comparing and analyzing a conventional examination and the quantitative light-induced fluorescence (QLF) technique. Selected were 297 teeth of 153 patients to take QLF images and bitewing radiographs. Occlusal dental caries, proximal dental caries and cracks were evaluated and scored using QLF, X-ray and/or visual criteria. The sensitivity, specificity, and area under the curve (AUC) of a receiver operating characteristic analysis were calculated. Two fluorescence parameters (|ΔFmax| and ΔRmax) were utilized to evaluate the fluorescence pattern according to the severity of lesions based on QLF or X-ray criteria. QLF showed higher scores for detecting occlusal dental caries and cracks than the conventional method. ΔRmax increased more clearly than ΔFmax did with occlusal dental caries. The |ΔFmax| values of occlusal dental caries, proximal dental caries and cracks showed good AUC levels (0.84, 0.81 and 0.83, respectively). The ΔRmax of occlusal dental caries showed the highest AUC (0.91) and the ΔRmax of proximal dental caries showed a fail level (0.59) compared to bitewing radiographs. The QLF image could visualize and estimate the degree of occlusal dental caries or cracks. Consequently, the QLF technique may be an adjunct tool to conventional methods for the detection of occlusal caries and peripheral cracks.

In this paper, we propose several methods to improve the performance of multiple object tracking (MOT), especially for humans, in dynamic environments such as robots and autonomous vehicles. The first method is to restore and re-detect unreliable results to improve the detection. The second is to restore noisy regions in the image before the tracking association to improve the identification. To implement the image restoration function used in these two methods, an image inference model based on SRGAN (super-resolution generative adversarial networks) is used. Finally, the third method includes an association method using face features to reduce failures in the tracking association. Three distance measurements are designed so that this method can be applied to various environments. In order to validate the effectiveness of our proposed methods, we select two baseline trackers for comparative experiments and construct a robotic environment that interacts with real people and provides services. Experimental results demonstrate that the proposed methods efficiently overcome dynamic situations and show favorable performance in general situations.

Zero-shot recognition (ZSR) aims to perform visual classification by category in the absence of training samples. The focus in most traditional ZSR models is using semantic knowledge about familiar categories to represent unfamiliar categories with only the visual appearance of an unseen object. In this research, we consider not only visual information but context to enhance the classifier’s cognitive ability in a multi-object scene. We propose a novel method, contextual inference, that uses external resources such as knowledge graphs and semantic embedding spaces to obtain similarity measures between an unseen object and its surrounding objects. Using the intuition that close contexts involve more related associations than distant ones, distance weighting is applied to each piece of surrounding information with a newly defined distance calculation formula. We integrated contextual inference into traditional ZSR models to calibrate their visual predictions, and performed extensive experiments on two different datasets for comparative evaluations. The experimental results demonstrate the effectiveness of our method through significant enhancements in performance.

Recently, several studies have focused on image-to-image translation. However, the quality of the translation results is lacking in certain respects. We propose a new image-to-image translation method to minimize such shortcomings using an auto-encoder and an auto-decoder. This method includes pre-training two auto-encoders and decoder pairs for each source and target image domain, cross-connecting two pairs and adding a feature mapping layer. Our method is quite simple and straightforward to adopt but very effective in practice, and we experimentally demonstrated that our method can significantly enhance the quality of image-to-image translation. We used the well-known cityscapes, horse2zebra, cat2dog, maps, summer2winter, and night2day datasets. Our method shows qualitative and quantitative improvements over existing models.

Classifying semantic relations between entity pairs in sentences is an important task in natural language processing (NLP). Most previous models applied to relation classification rely on high-level lexical and syntactic features obtained by NLP tools such as WordNet, the dependency parser, part-of-speech (POS) tagger, and named entity recognizers (NER). In addition, state-of-the-art neural models based on attention mechanisms do not fully utilize information related to the entity, which may be the most crucial feature for relation classification. To address these issues, we propose a novel end-to-end recurrent neural model that incorporates an entity-aware attention mechanism with a latent entity typing (LET) method. Our model not only effectively utilizes entities and their latent types as features, but also builds word representations by applying self-attention based on symmetrical similarity of a sentence itself. Moreover, the model is interpretable by visualizing applied attention mechanisms. Experimental results obtained with the SemEval-2010 Task 8 dataset, which is one of the most popular relation classification tasks, demonstrate that our model outperforms existing state-of-the-art models without any high-level features.

Recommendation systems are widely used in conjunction with many popular personalized services, which enables people to find not only content items they are currently interested in, but also those in which they might become interested. Many recommendation systems employ the memory-based collaborative filtering (CF) method, which has been generally accepted as one of consensus approaches. Despite the usefulness of the CF method for successful recommendation, several limitations remain, such as sparsity and cold-start problems that degrade the performance of CF systems in practice. To overcome these limitations, we propose a content-metadata-based approach that uses content-metadata in an effective way. By complementarily combining content-metadata with conventional user-content ratings and trust network information, our proposed approach remarkably increases the amount of suggested content and accurately recommends a large number of additional content items. Experimental results show a significant enhancement of performance, especially under a sparse rating environment.

Much like humans focus solely on object movement to understand actions, directing a deep learning model's attention to the core contexts within videos is crucial for improving video comprehension. In the recent study, VideoMasked Auto-Encoder (VideoMAE) employs a pre-training approach with a high ratio of tube masking and reconstruction, effectively mitigating spatial bias due to temporal redundancy in full video frames. This steers the model's focus toward detailed temporal contexts. However, as the VideoMAE still relies on full video frames during the action recognition stage, it may exhibit a progressive shift in attention towards spatial contexts, deteriorating its ability to capture the main spatio-temporal contexts. To address this issue, we propose an attention-directing module named Transformer Encoder Attention Module (TEAM). This proposed module effectively directs the model's attention to the core characteristics within each video, inherently mitigating spatial bias. The TEAM first figures out the core features among the overall extracted features fromeach video. After that, it discerns the specific parts of the video where those features are located, encouraging themodel to focusmore on these informative parts. Consequently, during the action recognition stage, the proposed TEAM effectively shifts the VideoMAE's attention from spatial contexts towards the core spatio-temporal contexts. This attention-shift manner alleviates the spatial bias in the model and simultaneously enhances its ability to capture precise video contexts. We conduct extensive experiments to explore the optimal configuration that enables the TEAM to fulfill its intended design purpose and facilitates its seamless integrationwith theVideoMAE framework. The integratedmodel, i.e., VideoMAE+TEAM, outperforms the existing VideoMAE by a significant margin on Something-Something-V2 (71.3% vs. 70.3%). Moreover, the qualitative comparisons demonstrate that theTEAMencourages themodel to disregard insignificant features and focus more on the essential video features, capturing more detailed spatio-temporal contexts within the video. [ABSTRACT FROM AUTHOR]