Replay-based class incremental learning (CIL) mitigates catastrophic forgetting by retaining limited historical data. However, the class imbalance between historical and new class data prevents catastrophic forgetting from being fully resolved. Existing approaches rely solely on buffer data to approximate the historical distribution, neglecting the incomplete historical distribution modeling caused by feature drift in the evolving feature space, limiting their effectiveness. To address this challenge, we propose a novel Feature Drift oriented Distribution Reconstruction (FDDR) framework, which reconstructs complete historical distributions utilizing both historical and new knowledge. Specifically, FDDR mitigates the structure degradation of historical distributions caused by feature drift by aligning distributions between the current and previous feature spaces, reserving sufficient feature spaces for subsequent distribution reconstruction. Based on the preserved space, FDDR supplements the sparse historical distribution by incorporating weighted new-class distributions according to inter-class relationships, which are derived from joint prototype cosine similarities. In particular, each joint prototype is constructed from both evolved historical prototypes and buffer prototypes, thereby yielding a more reliable similarity measure. Finally, building upon the reconstructed distributions, FDDR generates pseudo-features to correct biased classifier and further fine-tunes the feature space to achieve superior feature space reconstruction. Comprehensive experiments on widely adopted CIL benchmarks verify the effectiveness of FDDR. Code is available at https://github.com/njustkmg/ECAI25-FDDR.

Incremental learning (IL) aims to sequentially learn new tasks while mitigating catastrophic forgetting. Among various IL strategies, parameter-isolation methods stand out by using mask techniques to allocate distinct parameters to each task, explicitly addressing forgetting. However, existing approaches often disregard parameter dependencies, resulting in an over-reliance on newly allocated parameters. To address this issue, we propose Probabilistic Group Mask selection (PGM), a group-wise approach that captures parameter dependencies by exploring candidate masks within each group. Specifically, PGM partitions parameters into groups with multiple candidate masks, assigning probabilities to these masks and leveraging Gumbel-Softmax for differentiable sampling, enabling efficient optimization of the discrete mask selection process. Our theoretical analysis demonstrates that incorporating parameter dependencies enhances sub-network selection. Experiments conducted on standard benchmarks confirm its superior effectiveness compared to existing IL approaches. The source code is available at: \url{https://github.com/njustkmg/ICML25-PGM}.

Vision-language retrieval aims to perform cross-modal instances search by learning consistent vision-language representations. However, in real applications, vision-language divergence always results in strong and weak modalities and different modalities have various performances in uni-modal tasks. In this paper, we reveal that traditional vision-language hard consistency disrupts the relationships among uni-modal instances considering the weak-strong modal scenario, causing a decline in uni-modal retrieval capability. To address this issue, we propose Coordinated Vision-Language Retrieval (CoVLR), a solution that cooperatively optimizes both cross-modal consistency and intra-modal structure-preserving objectives via a meta-learning strategy. Specifically, CoVLR utilizes intra-modal structure-preserving as the meta-test task to validate the cross-modal consistency loss, which is considered the meta-train task. The effectiveness of CoVLR is validated through extensive experiments on commonly used datasets, which demonstrate superior results compared to other baselines on various retrieval scenarios.

Multimodal learning falls into the trap of the optimization dilemma due to the modality imbalance phenomenon, leading to unsatisfactory performance in real applications. A core reason for modality imbalance is that the models of each modality converge at different rates. Many attempts naturally focus on adjusting learning procedures adaptively. Essentially, the reason why models converge at different rates is because the difficulty of fitting category labels is inconsistent for each modality during learning. From the perspective of fitting labels, we find that appropriate positive intervention label fitting can correct this difference in learning ability. By exploiting the ability of contrastive learning to intervene in the learning of category label fitting, we propose a novel multimodal learning approach that dynamically integrates unsupervised contrastive learning and supervised multimodal learning to address the modality imbalance problem. We find that a simple yet heuristic integration strategy can significantly alleviate the modality imbalance phenomenon. Moreover, we design a learning-based integration strategy to integrate two losses dynamically, further improving the performance. Experiments on widely used datasets demonstrate the superiority of our method compared with state-of-the-art (SOTA) multimodal learning approaches. The code is available at https://github.com/njustkmg/NeurIPS24-LFM.