With the rise of powerful neural networks and large-scale datasets, pre-trained models (PTMs) exhibiting strong generalization capabilities have been successfully applied to class incremental learning (CIL) tasks. It is widely believed the key factors are generalizability and adaptability of PTMs, a better tradeoff leads to better performance. In this work, we dive into dissecting the representation of PTMs in CIL and try to answer the following question: What makes PTMs work in CIL and how can it be taken to the next level? Surprisingly, we find that PTMs with different basic architectures commonly show a strong bias towards texture. Based on that, we put forward a hypothesis that the bottlenecks of PTMs in CIL are derived from the drawbacks of inherent generalizability after pretraining. This texture bias limits the representation capability of PTMs, thereby affecting the performance on CIL tasks. To address this challenge, we first propose a novel paradigm that fundamentally rethinks the dual-branch architecture through strategic integration of a PTM with strong contour bias, effectively superseding conventional PTMs with improved adaptability. Empirical results demonstrate that, within the framework of our proposed paradigm, the dual-branch network achieves substantially improved performance compared to conventional approaches that emphasize balancing generalizability and adaptability. Finally, we advance the validity of our paradigm through a new method, Bi-branch network with Contour-enhanced features (BiCon), which further strengthens the contour perception bias of PTMs. Extensive experiments across six diverse datasets consistently demonstrate the superiority of our method.

Class imbalance significantly hampers recognition accuracy in image classification, which is particularly prevalent in special field datasets. Traditional methodologies addressing imbalance, including techniques like oversampling, have shown limited effectiveness. Recently, the advanced generative capabilities of diffusion models garnered attention, with subsequent developments like SinDiffusion enabling new possibilities for rare class sample generation. However, the outputs of these models often significantly deviate from the original targets. Building on this foundation, our approach incorporates edge constraints to ensure newly generated samples not only differ from originals but also retain essential target features. Experiments validate the effectiveness of our method, highlighting the potential to mitigate class imbalance challenges in classification tasks.

Besides data sparsity and cold start, recommender systems often face the problems of selection bias and exposure bias. These problems influence the accuracy of recommendations and easily lead to overrecommendations. This paper proposes a recommendation approach based on heterogeneous network and dynamic knowledge graph (HN-DKG). The main steps include (1) determining the implicit preferences of users according to user’s cross-domain and cross-platform behaviors to form multimodal nodes and then building a heterogeneous knowledge graph; (2) Applying an improved multihead attention mechanism of the graph attention network (GAT) to realize the relationship enhancement of multimodal nodes and constructing a dynamic knowledge graph; and (3) Leveraging RippleNet to discover user’s layered potential interests and rating candidate items. In which, some mechanisms, such as user seed clusters, propagation blocking, and random seed mechanisms, are designed to obtain more accurate and diverse recommendations. In this paper, the public datasets are used to evaluate the performance of algorithms, and the experimental results show that the proposed method has good performance in the effectiveness and diversity of recommendations. On the MovieLens-1M dataset, the proposed model is 18%, 9%, and 2% higher than KGAT on F1, NDCG@10, and AUC and 20%, 2%, and 0.9% higher than RippleNet, respectively. On the Amazon Book dataset, the proposed model is 12%, 3%, and 2.5% higher than NFM on F1, NDCG@10, and AUC and 0.8%, 2.3%, and 0.35% higher than RippleNet, respectively.

In recent years, there has been a prominent focus on enhancing the quality of features derived from convolutional neural networks (CNNs) within the field of polarimetric synthetic aperture radar (PolSAR) image classification. Targeting this challenge, this article first visualizes the lack of discriminability and generalizability in CNN features through several empirical observations. Subsequently, we explain why these problems arise from a causal perspective, accomplished by means of a structural causal model (SCM) constructed according to the training and testing process of CNNs. This SCM facilitates the identification of variables that affect the quality of PolSAR image feature learning, as well as an intervention on those variables using backdoor adjustment. Building upon this groundwork, a novel causal inference-guided feature enhancement framework is constructed. It can be seamlessly integrated into any CNN-based PolSAR image classifier in a plug-and-play manner, enabling the enhanced classifier to filter out interference information and prevent model overfitting. These two aspects bring better feature discriminability and generalizability, respectively, leading to improved classification performance. Experimental results on four widely-used PolSAR image datasets demonstrate the effectiveness of our proposed framework. We integrate it into several mainstream methods in the field and show that the accuracy of the enhanced classifier is improved compared to the original model.

Accurate forecasting of tropical cyclone (TC) plays a critical role in the prevention and defense of TC disasters. We must explore a more accurate method for TC prediction. Deep learning methods are increasingly being implemented to make TC prediction more accurate. However, most existing methods lack a generic framework for adapting heterogeneous meteorological data and do not focus on the importance of the environment. Therefore, we propose a Multi-Generator Tropical Cyclone Forecasting model (MGTCF), a generic, extensible, multi-modal TC prediction model with the key modules of Generator Chooser Network (GC-Net) and Environment Net (Env-Net). The proposed method can utilize heterogeneous meteorologic data efficiently and mine environmental factors. In addition, the Multi-generator with Generator Chooser Net is proposed to tackle the drawbacks of single-generator TC prediction methods: the prediction of undesired out-of-distribution samples and the problems stemming from insufficient learning ability. To prove the effectiveness of MGTCF, we conduct extensive experiments on the China Meteorological Administration Tropical Cyclone Best Track Dataset. MGTCF obtains better performance compared with other deep learning methods and outperforms the official prediction method of the China Central Meteorological Observatory in most indexes.

In hands-free communication system, the coupling between loudspeaker and microphone generates echo signal, which can severely influence the quality of communication.Meanwhile, various types of noise in communication environments further reduce speech quality and intelligibility.It is difficult to extract the near-end signal from the microphone signal within one step, especially in low signal-to-noise ratio scenarios.In this paper, we propose a deep complex network approach to address this issue.Specially, we decompose the stereophonic acoustic echo cancellation into two stages, including linear stereophonic acoustic echo cancellation module and residual echo suppression module, where both modules are based on deep learning architectures.A multi-frame filtering strategy is introduced to benefit the estimation of linear echo by capturing more interframe information.Moreover, we decouple the complex spectral mapping into magnitude estimation and complex spectrum refinement.Experimental results demonstrate that our proposed approach achieves stage-of-the-art performance over previous advanced algorithms under various conditions.