Recognizing distinct categories from the same super-category is a highly challenging task due to small inter-class and large intra-class variances. Despite the significant progress of deep convolutional neural network-based methods over the past few years, they are incapable of locating complete salient regions with respect to the object and parts (OPs) of an input image. Locating complete salient regions of OPs is very crucial to fine-grained visual categorization (FGVC), as it contributes to adequately capturing the salient features of OPs. Besides, most existing methods focus on extracting OPs’ features in a unidirectional way, neglecting potentially valuable information between OPs and that within themselves of different feature levels. Consequently, the quality of fine-grained features that are discriminative for FGVC is affected. To address these issues, herein an interaction feedback network with salient-region-aware augmentation (IFN-SA) is proposed. Specifically, a preliminary feature learning (PFL) module is first designed through the use of attention-based object and part streams, which preserve the contextual details semantically associated with OPs, facilitating the location of the complete salient regions and acquisition of the salient features of OPs. Following this, a progressive representation learning (PRL) module is built through the specially-designed vertical interaction mechanism (VIM) and horizontal feedback mechanism (HFM). VIM captures the high-level features via the interactions between OPs, while HFM calibrates the salient features with the feedbacks from the high-level ones and progressively excavates the latter ones. The experimental results especially the performance comparisons on four benchmark datasets validate the excellent performance of IFN-SA for FGVC.

Non-gradient descent learning schemes have recently gained attention for their efficient training and simplification of learning control parameter tuning. However, most methods are designed for tabular data (or image data represented in the form of 1D feature vectors), which limits their ability to capture the spatial and semantic richness of structured data. Additionally, to enhance representation learning capacity, these models often struggle to accurately determine the required complexity before construction, leading to unnecessary model complexity. To address these issues, this paper proposes a Synergic Learning System based on Swarm of Non-Gradient Learners (SONG). The system introduces a neural network-based model composed of lightweight, efficient, and scalable learning units, including Pseudoinverse Learners (PILers), a learnable Patch Respective Field (PRF), and a Pseudoinverse Learning-based Autoencoder (PILAE) based MLP, enhancing representation learning for both tabular and image data. Inspired by synergetic learning system and emergence theory, SONG dynamically aggregates multiple base models, adaptively determining subsystem quantity and hierarchical depth. Experimental results show that SONG outperforms existing competitive baseline methods on several publicly available benchmark datasets, achieving efficient training, low hyperparameter sensitivity, and superior generalization, while generally exhibiting emergent phenomena.

Wind energy is an integral component of a solution to the challenges of global energy crises and environmental degradation, since it is a renewable and environmentally friendly source of electricity. However, the unpredictability of wind speed makes the efficient utilization of wind power challenging. To overcome this challenge, a hybrid wavelet decomposition-ensemble empirical mode decomposition-convolutional neural network-long short-term memory (WD-EEMD-CNN-LSTM) wind speed forecasting model based on WD, EEMD, CNN, and LSTM algorithms is proposed. First, WD is deployed to minimize the fluctuation in the wind speed, and then the wind signal is decomposed into eight IMFs and one residual. The top three IMFs are applied to the input of a proposed model and compared with other state-of-the-art single and simple hybrid models. Furthermore, the proposed model's outputs are incorporated into the traditional approach to formulating the dispatch problem, aiming to reduce the generation cost. The expected cost of power dispatch is calculated for 2-bus and IEEE 24 RTS systems. The results showed that the proposed model has the lowest cost for day-ahead and balance markets. Overall, the WD-EEMD-CNN-LSTM model demonstrates superior performance in wind speed forecasting accuracy, making it a promising solution for increasing wind power utilization.

While spiking neural networks (SNNs) have demonstrated remarkable efficiency in neuromorphic computing by emulating biological neuronal dynamics, their learning capabilities remain constrained by predominant focus on synaptic plasticity. This limitation overlooks critical neurobiological evidence showing that intrinsic neuronal plasticity, particularly dynamic threshold adaptation, plays an essential role in balancing neural responsiveness and signal fidelity. Inspired by two neurophysiological principles governing threshold regulation: 1) the inverse correlation between spiking thresholds and preceding depolarization rates, and 2) the proportional relationship between thresholds and average membrane potentials, we propose a Membrane Potential-Driven Adaptive Threshold Plasticity (MPD-ATP) framework. This biologically grounded mechanism establishes a dual-pathway control system where instantaneous depolarization rates and sustained membrane potential states jointly modulate neuronal thresholds through an adaptive scaling factor. The instantaneous depolarization rate dynamically lowers thresholds during strong input bursts, while the sustained average membrane potential adjusts the baseline threshold to stabilize firing during sparse input. This complementary regulation improves precision and robustness. Extensive evaluations on static (CIFAR-10/100) and neuromorphic (CIFAR10-DVS, DVSGesture) benchmarks demonstrate that MPD-ATP-enhanced networks achieve superior classification accuracy with enhanced noise robustness. Systematic ablation studies reveal that the coordinated interaction between depolarization-sensitive and membrane potential-proportional threshold adjustments is critical for preventing signal saturation in high-activity networks while mitigating under-activation in sparse-input scenarios.

Data-driven models are widely adopted, but their opaque nature often hinders understanding and trust. This paper introduces a novel human-in-the-loop framework for interactive and explainable data-driven modeling. Central to our approach is the integration of fuzzy neural networks (FNNs) with evolutionary optimization, guided by continuous human feedback. Unlike traditional methods, our framework enables users to actively participate, influencing the FNN’s decision-making in real-time. This feedback directly informs the model’s structural and parameter updates, resulting in models that are both dynamically adaptable and inherently transparent through their fuzzy logic. Experimental results demonstrate this framework significantly enhances model performance while fostering trust and comprehension by providing clear reasoning processes. This work advances human-centric data modeling, improving explainability and interactivity for complex data relationships.

State Space Models (SSMs) and Graph Neural Networks (GNNs) have been demonstrated to be highly effective in visual tasks. Existing vision SSMs capture features in a focused manner through a sequential scanning pattern, achieving near-linear complexity but constrained by limited directional receptive fields, which restrict image comprehension. In contrast, current vision GNNs’ flexible graph modeling captures global and complex relationships but tends to lose focus on key regions of the image due to their more sparse attention. To address these challenges, we propose a novel architecture, GMamba, which provides the ability to flexibly capture diverse features and keeps focus on the key regions within images. In GMamba, we design the Hybrid Adaptive Perceptor (HAP) to achieve focused and flexible visual perception through the adaptive integration of GNNs and SSMs. Extensive experiments across various tasks and datasets validate the superior performance of GMamba, and the in-depth analyses reveal its effectiveness in visual perception. These results highlight the promising potential of combining GNNs and SSMs as the visual backbone, providing a new insight for advancing capabilities in visual tasks.

The spread of misinformation on social platforms is amplified by multimodal posts and the discourse they trigger. We present <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LANTERN</monospace>, a lightweight framework that detects fake posts by pairing a compact LVLM (Large Visio-Language Model) with a graph-to-text rendering of reply threads. Comment trees are pruned breadth-first and serialized with depth-based indentation and in-line metadata (author, engagement), then concatenated with image tokens; the LVLM is fine-tuned with low-rank adapters and a linear head, using the end-of-sequence state for classification. On <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Fakeddit</i>, <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LANTERN</monospace> attains state-of-the-art accuracy with only 150 k training samples, achieving <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">97.16%</b> (2-way), <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">97.05%</b> (3-way), <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">93.72%</b> (6-way). To assess versatility beyond Reddit, we additionally evaluate <monospace xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LANTERN-TEXT</monospace> on <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PHEME</i> (text-only rumor cascades) and adapt the pipeline to <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Twitter15/16</i> (structure-only replies), observing robust gains from shallow discourse and and measurable gains even when reply text is absent. Large-scale multimodal pretraining plus lightweight adaptation and structured conversational context yields a practical and reproducible approach to misinformation detection.

The fusion of structural and functional brain network analysis has been widely applied in the analysis of various brain diseases, especially in identifying the progression stages of Alzheimer’s disease. However, most existing multimodal brain network analysis methods separately construct functional and structural networks, making it difficult to incorporate complementary information from several modalities. To tackle this difficulty, we propose AMFusion(Adaptive Multimodal Fusion), a unified brain network construction framework that jointly learns from functional and structural images, thereby efficiently resolving connectivity and node feature learning problems. Our approach begins with a designed quantization encoder to extract structural features from DTI, while dynamic functional connectivity (FC) is constructed from fMRI. Subsequently, a multilevel brain network fusion module is employed to learn and integrate brain connections, considering interactions across different temporal and spatial scales. Finally, a classifier guides further optimization of the brain network, facilitating multi-class disease classification. The performance of the proposed AMFusion was validated using the authentic dataset from ADNI, and experimental results demonstrated the effectiveness of AMFusion. By considering both connection patterns and node features, our method overcomes the limitations of existing approaches, providing a more comprehensive and robust framework for analyzing and diagnosing AD.

The goal of medical image translation tasks is to design models that transform source domain images to target domain images, with the potential to significantly enhance diagnostic accuracy and reduce patient radiation exposure. Current advanced methods employ generative models to characterize the conditional distribution of target domain images, typically incorporating high frequency edge information from source domain images as generation conditions to maintain anatomical consistency between the two domains. However, directly introducing edge information may introduce noise and misleading information, adversely affecting the generative model. To address this issue, we propose a Fast and Slow Diffusion framework with Fuzzy Contextual Representation. Specifically, High-frequency edge images undergo fast diffusion while original images follow a slow diffusion process, coupled by a contextual pool that dynamically stores edge patterns from the fast path. Furthermore, we apply fuzzification to the contextual pool to enhance its expressiveness and use the resulting fuzzy representations to guide the reverse process of the slow diffusion, ensuring the accurate preservation of edge structures. Comprehensive experiments demonstrate that our proposed framework achieves performance comparable to existing state-of-the-art methods in medical image translation tasks. In addition, quantitative experiments demonstrate that fuzzified edge representations enhance generative models' capability in precisely perceiving and retaining high-frequency components.

Recently, image retrieval from massive images has been considered as a crucial research topic. Current deep hashing methods mostly rely on convolutional neural network (CNN) as the backbone network. However, the CNN-based deep hashing methods require multiple convolution operations to capture global information, which lead to the loss of many shallow-level features in the process of feature extraction and affect the accuracy of binary hashing codes. To address such limitations, a new deep hashing method is proposed and studied in this paper by adopting the vision transformer model as the backbone network and designing a spatial reduction layer with channel attention mechanism fused. Such a deep hashing method can capture long-range dependencies effectively. The experiments on the CIFAR-10, NUSWIDE and ImageNet datasets indicate that the proposed method achieves higher mean average precision than other state-of-the-art hashing methods, thereby confirming the outstanding performance on image retrieval.