Zicheng Pan

Few-shot class-incremental learning (FSCIL) aims to learn from a sequence of incremental data sessions with a limited number of samples in each class. The main issues it encounters are the risk of forgetting previously learned data when introducing new data classes, as well as not being able to adapt the old model to new data due to limited training samples. Existing state-of-the-art solutions normally utilize pre-trained models with fixed backbone parameters to avoid forgetting old knowledge. While this strategy preserves previously learned features, the fixed nature of the backbone limits the model’s ability to learn optimal representations for unseen classes, which compromises performance on new class increments. In this paper, we propose a novel SEssion-Guided Attention framework (SEGA) to tackle this challenge. SEGA exploits the class relationships within each incremental session by assessing how test samples relate to class prototypes. This allows accurate incremental session identification for test data, leading to more precise classifications. In addition, an attention module is introduced for each incremental session to further utilize the feature from the fixed backbone. As the session of the testing image is determined, we can fine-tune the feature with the corresponding attention module to better cluster the sample within the selected session. Our approach adopts the fixed backbone strategy to avoid forgetting the old knowledge while achieving novel data adaptation. Experimental results on three FSCIL datasets consistently demonstrate the superior adaptability of the proposed SEGA framework in FSCIL tasks.

Source-free domain adaptation (SFDA) aims to transfer knowledge from a labelled source domain to an unlabelled target domain. A major challenge in SFDA is deriving accurate categorical information for the target domain, especially when sample embeddings from different classes appear similar. This issue is particularly pronounced in fine-grained visual categorization tasks, where inter-class differences are subtle. To overcome this challenge, we introduce a novel ETF-Informed Attention Network (EIANet) to separate class prototypes by utilizing attention and neural collapse principles. More specifically, EIANet employs a simplex Equiangular Tight Frame (ETF) classifier in conjunction with an attention mechanism, facilitating the model to focus on discriminative features and ensuring maximum class prototype separation. This innovative approach effectively enlarges the feature difference between different classes in the latent space by locating salient regions, thereby preventing the misclassification of similar but distinct category samples and providing more accurate categorical information to guide the fine-tuning process on the target domain. Experimental results across four SFDA datasets validate EIANet's state-of-the-art performance. Code is available at: https://github.com/zichengpan/EIANet.

Few-shot class-incremental learning (FSCIL) aims to continually learn novel data with limited samples. One of the major challenges is the catastrophic forgetting problem of old knowledge while training the model on new data. To alleviate this problem, recent state-of-the-art methods adopt a well-trained static network with fixed parameters at incremental learning stages to maintain old knowledge. These methods suffer from the poor adaptation of the old model with new knowledge. In this work, a dynamic clustering and recovering network (DyCR) is proposed to tackle the adaptation problem and effectively mitigate the forgetting phenomena on FSCIL tasks. Unlike static FSCIL methods, the proposed DyCR network is dynamic and trainable during the incremental learning stages, which makes the network capable of learning new features and better adapting to novel data. To address the forgetting problem and improve the model performance, a novel orthogonal decomposition mechanism is developed to split the feature embeddings into context and category information. The context part is preserved and utilized to recover old class features in future incremental learning stages, which can mitigate the forgetting problem with a much smaller size of data than saving the raw exemplars. The category part is used to optimize the feature embedding space by moving different classes of samples far apart and squeezing the sample distances within the same classes during the training stage. Experiments show that the DyCR network outperforms existing methods on four benchmark datasets. The code is available at: https://github.com/zichengpan/DyCR.

Learning discriminative representation with limited training samples is emerging as an important yet challenging visual categorization task. While prior work has shown that incorporating self-supervised learning can improve performance, we found that the direct use of canonical metric in a Lie group is theoretically incorrect. In this article, we prove that a valid optimization measurement should be a canonical metric on Lie algebra. Based on the theoretical finding, this article introduces a novel self-supervised Lie algebra network (SLA-Net) representation learning framework. Via minimizing canonical metric distance between target and predicted Lie algebra representation within a computationally convenient vector space, SLA-Net avoids computing nontrivial geodesic (locally length-minimizing curve) metric on a manifold (curved space). By simultaneously optimizing a single set of parameters shared by self-supervised learning and supervised classification, the proposed SLA-Net gains improved generalization capability. Comprehensive evaluation results on eight public datasets show the effectiveness of SLA-Net for visual categorization with limited samples.