Retrieval-enhanced adversarial training with dynamic memory-augmented attention for image paragraph captioning

Evaluation metrics play an important role in accessing the performance of a regression method. In practice, these multiple evaluation metrics can be used in two ways. The first way defines a loss function by aggregating multiple metrics, while the second way defines a multiobjective loss function by considering each metric as an objective function. In this paper, we propose a new way to use multiple evaluation metrics, which is different from the aggregating method and the mutliobjective method. Our method is based on genetic programming. The idea is to randomly use one metric in each iteration of the selection operator. Therefore, multiple metrics can be used alternatively in the running process. To validate the effectiveness of our new approach, we conduct experiments on ten benchmark datasets. The experimental results show that the new approach can improve the population diversity, and can achieve the performance better than or similar to that of the traditional symbolic regression algorithms.

There are more than 2,000 listed companies on the UK’s London Stock Exchange, divided into 11 sectors who are required to communicate their financial results at least twice in a single financial year. UK annual reports are very lengthy documents with around 80 pages on average. In this study, we aim to benchmark a variety of summarisation methods on a set of different pre-trained transformers with different extraction techniques. In addition, we considered multiple evaluation metrics in order to investigate their differing behaviour and applicability on a dataset from the Financial Narrative Summarisation (FNS 2020) shared task, which is composed of annual reports published by firms listed on the London Stock Exchange and their corresponding summaries. We hypothesise that some evaluation metrics do not reflect true summarisation ability and propose a novel BRUGEscore metric, as the harmonic mean of ROUGE-2 and BERTscore. Finally, we perform a statistical significance test on our results to verify whether they are statistically robust, alongside an adversarial analysis task with three different corruption methods.

Biomedical knowledge graphs (KGs), which can help with the understanding of complex biological systems and pathologies, have begun to play a critical role in medical practice and research. However, challenges remain in their embedding and use due to their complex nature and the specific demands of their construction. Existing studies often suffer from problems such as sparse and noisy datasets, insufficient modeling methods and non-uniform evaluation metrics. In this work, we established a comprehensive KG system for the biomedical field in an attempt to bridge the gap. Here, we introduced PharmKG, a multi-relational, attributed biomedical KG, composed of more than 500000 individual interconnections between genes, drugs and diseases, with 29 relation types over a vocabulary of ~8000 disambiguated entities. Each entity in PharmKG is attached with heterogeneous, domain-specific information obtained from multi-omics data, i.e. gene expression, chemical structure and disease word embedding, while preserving the semantic and biomedical features. For baselines, we offered nine state-of-the-art KG embedding (KGE) approaches and a new biological, intuitive, graph neural network-based KGE method that uses a combination of both global network structure and heterogeneous domain features. Based on the proposed benchmark, we conducted extensive experiments to assess these KGE models using multiple evaluation metrics. Finally, we discussed our observations across various downstream biological tasks and provide insights and guidelines for how to use a KG in biomedicine. We hope that the unprecedented quality and diversity of PharmKG will lead to advances in biomedical KG construction, embedding and application.

Multi-modal knowledge graph completion (MMKGC) leverages visual, textual, and structural information to predict missing entities in knowledge graphs. However, existing approaches predominantly rely on early fusion strategies that project heterogeneous modalities into unified representations, diminishing modality-specific discriminative capabilities and failing to fully exploit inter-modal complementarity. Unlike prior fusion-based models that merge all modalities into a single latent space, our approach treats each modality as an independent reasoning expert and integrates their complementary outputs through an interpretable ensemble mechanism. To address these limitations, we propose the Multi-Branch Ensemble with Cross-Modal Fusion (MBE-CF) framework, which preserves modality-specific characteristics through independent reasoning pathways while capturing cross-modal interactions via dedicated fusion mechanisms. Our approach introduces three key innovations: (1) a multi-modal attention fusion (MMAF) module that employs multi-head self-attention mechanisms to capture cross-modal interactions and obtain robust fused multi-modal representations, (2) a context triple encoder (CTE) that processes head-relation-tail sequences through modality-specific Transformer encoders to inject relational context into entity representations, and (3) an Ensemble Inference (EI) approach that uses independent Tucker decoders for structural, visual, textual, and multi-modal branches, aggregates per-branch scores at inference, and optimizes during training a Decision Fusion objective that sums branch-specific head and tail losses. The framework maintains architectural symmetry across modalities while facilitating comprehensive integration of heterogeneous reasoning patterns to leverage the complementary strengths of different information sources. Extensive experiments on three benchmark datasets (DB15K, MKG-W, and MKG-Y) demonstrate that MBE-CF achieves state-of-the-art performance, outperforming 9 baseline methods across multiple evaluation metrics.

Temporal knowledge graph (TKG) reasoning has attracted significant attention. Recent approaches for modeling historical information have led to great advances. However, the problems of time variability and unseen entities have become two major obstacles preventing further development. The time variability problem means that different historical timestamps play different roles in the inference process. Furthermore, in the context of time variability, the unseen entity problem means that a query cannot obtain a predicted entity that is unseen in the scale-varying history rather than in a fixed set, thus turning from static to dynamic. In this paper, we propose a novel method named DHU-NET for addressing the time variability challenge and the dynamic unseen entity challenge derived from it. With regard to the former concern, we propose a time-distributed representation learning method based on a graph convolutional network(GCN) and a self-attention mechanism, which learns the distributed representations of facts at different historical timestamps and comprehensively pays different levels of attention to the different timestamps. With regard to the latter issue, we extract the unseen entities from a global static KG based on a copy mechanism and bring them into consideration during the final prediction step. Experiments on six benchmark datasets demonstrate the substantial improvements achieved by DHUNET in terms of multiple evaluation metrics. Our released codes are available at https://github.com/CGCL-codes/DHUNET.

Local positive and negative correlation-based k-labelsets for multi-label classification

RGB-D salient object detection (RGB-D SOD) aims to accurately localize and segment visually salient objects by jointly leveraging RGB images and depth maps. Some existing methods rely on static fusion strategies with fixed paths and weights, which treat all regions equally and fail to capture the varying importance of different regions and modalities. Although some attention-based methods alleviate the limitations of static fusion by assigning adaptive weights to different regions and modalities, the quality of RGB and depth data may degrade in real-world scenarios due to sensor noise, illumination changes, or environmental interference. These attention-based methods often overlook inter-modality quality differences and complementarity, making them prone to over-relying on a certain modality, which can lead to noise introduction, feature conflicts, and performance degradation. To address these limitations, this paper proposes a novel dynamic feature routing and fusion framework for RGB-D SOD, which adaptively adjusts the fusion strategy according to the quality of input modalities. To enable modality quality awareness, the proposed method characterizes the modality complementarity between RGB and depth features in a task-driven manner inspired by information-theoretic principles. We introduce a task-relevance scoring function which is integrated with a mutual information estimator to quantify such complementarity, and emphasizes task-relevant features while suppressing redundancy. A dynamic routing module is then designed to perform feature selection guided by the captured complementarity. In addition, we propose a novel cross-modal fusion module to adaptively fuse the features selected by the dynamic routing module, which effectively enhances complementary representations while suppressing redundant features and noise interference. Extensive experiments conducted on seven public RGB-D SOD benchmark datasets demonstrate that the proposed method consistently achieves competitive performance, outperforming existing methods by an average of approximately 1% across multiple evaluation metrics. Notably, in challenging scenarios with severe modality quality degradation, the proposed method outperforms existing best-performing methods by up to 1.8%, demonstrating strong robustness against cluttered backgrounds, complex object structures, and diverse object scales. Overall, the proposed dynamic fusion framework provides a novel solution to modality quality imbalance in RGB-D salient object detection.

Dynamic acoustic compensation and adaptive focal training for personalized speech enhancement

Recently, multi-channel speech enhancement has drawn much interest due to the use of spatial information to distinguish target speech from interfering signal. To make full use of spatial information and neural network based masking estimation, we propose a multi-channel denoising neural network - Spatial DCCRN. Firstly, we extend S-DCCRN to multi -channel scenario, aiming at performing cascaded sub-channel and full-channel processing strategy, which can model different channels separately. Moreover, instead of only adopting multi-channel spectrum or concatenating first-channel's magnitude and IPD as the model's inputs, we apply an angle feature extraction module (AFE) to extract frame-level angle feature embeddings, which can help the model to apparently perceive spatial information. Finally, since the phenomenon of residual noise will be more serious when the noise and speech exist in the same time frequency (TF) bin, we particularly design a masking and mapping filtering method to substitute the traditional filter-and-sum operation, with the purpose of cascading coarsely denoising, dereverberation and residual noise suppression. The proposed model, Spatial-DCCRN, has surpassed EaBNet, FasNet as well as several competitive models on the L3DAS22 Challenge dataset. Not only the 3D scenario, Spatial-DCCRN outperforms state-of-the-art (SOTA) model MIMO-UNet by a large margin in multiple evaluation metrics on the multi-channel ConferencingSpeech2021 Challenge dataset. Ablation studies also demonstrate the effectiveness of different contributions.

An advanced deep neural network for fundus image analysis and enhancing diabetic retinopathy detection

Automated vehicles promise a future where drivers can engage in non-driving tasks without hands on the steering wheels for a prolonged period. Nevertheless, automated vehicles may still need to occasionally hand the control back to drivers due to technology limitations and legal requirements. While some systems determine the need for driver takeover using driver context and road condition to initiate a takeover request, studies show that the driver may not react to it. We present DeepTake, a novel deep neural network-based framework that predicts multiple aspects of takeover behavior to ensure that the driver is able to safely take over the control when engaged in non-driving tasks. Using features from vehicle data, driver biometrics, and subjective measurements, DeepTake predicts the driver's intention, time, and quality of takeover. We evaluate DeepTake performance using multiple evaluation metrics. Results show that DeepTake reliably predicts the takeover intention, time, and quality, with an accuracy of 96%, 93%, and 83%, respectively. Results also indicate that DeepTake outperforms previous state-of-the-art methods on predicting driver takeover time and quality. Our findings have implications for the algorithm development of driver monitoring and state detection.

Traditional community detection methods in attributed networks (eg, social network) usually disregard abundant node attribute information and only focus on structural information of a graph. Existing community detection methods in attributed networks are mostly applied in the detection of nonoverlapping communities and cannot be directly used to detect the overlapping structures. This article proposes an overlapping community detection algorithm in attributed networks. First, we employ the modified X‐means algorithm to cluster attributes to form different themes. Second, we employ the label propagation algorithm (LPA), which is based on neighborhood network conductance for priority and the rule of theme weight, to detect communities in each theme. Finally, we perform redundant processing to form the final community division. The proposed algorithm improves the X‐means algorithm to avoid the effects of outliers. Problems of LPA such as instability of division and adjacent communities being easily merged can be corrected by prioritizing the node neighborhood network conductance. As the community is detected in the attribute subspace, the algorithm can find overlapping communities. Experimental results on real‐attributed and synthetic‐attributed networks show that the performance of the proposed algorithm is excellent with multiple evaluation metrics.

Demand and supply gap analysis of Chinese new energy vehicle charging infrastructure: Based on CNN-LSTM prediction model

Existing state-of-the-art 3D instance segmentation methods perform semantic segmentation followed by grouping. The hard predictions are made when performing semantic segmentation such that each point is associated with a single class. However, the errors stemming from hard decision propagate into grouping that results in (1) low overlaps between the predicted instance with the ground truth and (2) substantial false positives. To address the aforementioned problems, this paper proposes a 3D instance segmentation method referred to as SoftGroup by performing bottom-up soft grouping followed by top-down refinement. SoftGroup allows each point to be associated with multiple classes to mitigate the problems stemming from semantic prediction errors and suppresses false positive instances by learning to categorize them as background. Experimental results on different datasets and multiple evaluation metrics demonstrate the efficacy of SoftGroup. Its performance surpasses the strongest prior method by a significant margin of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$+6.2\%$</tex> on the ScanNet v2 hidden test set and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$+6.8\%$</tex> on S3DIS Area 5 in terms of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$AP_{50}$</tex> . Soft-Group is also fast, running at 345ms per scan with a sin-gle Titan X on ScanNet v2 dataset. The source code and trained models for both datasets are available at https://github.com/thangvubk/SoftGroup.git.

Postprocessing forecast precipitation fields from numerical weather prediction models aims to produce ensemble forecasts that are of high quality at each grid cell and, importantly, are spatially structured in an appropriate manner. A conventional approach, the gridcell-by-gridcell postprocessing, typically consists of two steps: 1) perform statistical calibration separately at individual grid cells to generate unbiased, skillful, and reliable ensemble forecasts and 2) employ ensemble reordering to link ensemble members of all grid cells according to certain templates to form spatially structured ensemble forecasts. However, ensemble reordering techniques are generally problematic in practical use. For example, the well-known Schaake shuffle is often criticized for not considering real physical atmospheric conditions. In this context, a fundamentally new approach, namely, spatial-mode-based calibration (SMoC), has recently been developed for postprocessing forecast precipitation fields with inbuilt spatial structures, thereby eliminating the need for ensemble reordering. SMoC was tested on 1-day-ahead forecasts of heavy precipitation events and was found to produce ensemble forecasts with appropriate spatial structures. In this paper, we extend SMoC to calibrate forecasts of light and no precipitation events and forecasts at long lead times. We also compare SMoC with the gridcell-by-gridcell postprocessing. Results based on multiple evaluation metrics show that SMoC performs well in calibrating both forecasts of light and no precipitation events and forecasts at long lead times. Compared with the gridcell-by-gridcell postprocessing, SMoC produces ensemble forecasts with similar forecast skill, improved forecast reliability, and clearly better spatial structures. In addition, SMoC is computationally far more efficient.