91影视

In recent news, the prestigious international conference in the field of artificial intelligence and machine learning, Conference on Neural Information Processing Systems (NeurIPS), has announced the paper acceptance results for NeurIPS 2025. The “MSU-BIT-SMBU Joint Research Center of Applied Mathematics” led by Professor Li Chun has achieved remarkable success, with two papers from the team being accepted for presentation at the conference.

This year, the conference received 21,575 valid submissions, with 5,290 papers being accepted, resulting in an acceptance rate of 24.52%. As an A-level conference recognized by the China Computer Federation (CCF), NeurIPS is one of the top three international conferences in the fields of artificial intelligence and machine learning, alongside ICML and ICLR, making it highly influential on the global stage. NeurIPS maintains a rigorous paper selection process, emphasizing both theoretical and engineering innovations and their contributions to the field. The acceptance of two papers from the SMBU team highlights the university's solid research foundation and cutting-edge capabilities in artificial intelligence and interdisciplinary studies.

Both papers, with Professor Li Chun as the corresponding author, were co-authored by graduate students Qiu Xihang and Zhang Wanpeng (the latter as part of their undergraduate thesis), focusing on frontier topics such as emotion recognition under incomplete modalities and high-dimensional generative modeling. The papers cover various hot topics, including federated learning and AIGC image generation.

Paper: “Federated Dialogue-Semantic Diffusion for Emotion Recognition under Incomplete Modalities”. Authors: Qiu Xihang, Cheng Jiarong, Fang Yuhao, Zhang Wanpeng, Lu Yao, Zhang Ye, and Li Chun.

Multimodal Emotion Recognition in Conversation (MERC) enhances emotional understanding by integrating multiple modal signals. However, the unpredictable loss of modalities in real-world scenarios can significantly degrade the performance of existing methods. Traditional approaches rely on complete multi-modal data for training, often encountering issues such as semantic distortion when specific modalities are missing. The proposed framework, FedDISC, integrates federated learning into the task of modality completion for the first time. By aggregating models trained on client-specific data and broadcasting them to clients with missing modalities, FedDISC overcomes the dependency on complete modalities in a single client. Additionally, the DISC diffusion module captures conversational dependencies using the Dialogue Graph Network (DGN) and enhances semantic alignment via the Semantic Condition Network (SCN), ensuring the consistency of the completed modalities with the existing ones at the context, speaker identity, and semantic levels. We further propose an alternating freezing aggregation strategy, which promotes collaborative optimization by cyclically freezing both the completion module and the classifier module. Extensive experiments on the IEMOCAP, CMUMOSI, and CMUMOSEI datasets demonstrate that FedDISC achieves exceptional emotion classification performance under various modality loss scenarios, significantly outperforming existing methods.

Figure 1 Framework of DGN and SCN.

DGN captures context and speaker dependencies through graph networks, while SCN uses attention mechanisms to capture cross-modal semantic information.

Figure 2 t-SNE visualizations comparing the modality recovery performance of different methods when only one modality is available.

Compared to other methods, the features generated by FedDISC exhibit higher distribution similarity with the original modality features, demonstrating its effectiveness.

Table 1 Comparison results on the CMUMOSI and CMUMOSEI datasets with varying loss rates.

Paper: “Proper H?lder-Kullback Dirichlet Diffusion: A Framework for High Dimensional Generative Modeling” . Authors: Zhang Wanpeng, Fang Yuhao, Qiu Xihang, Cheng Jiarong, Hong Jialong, Zhai Bin, Zhou Qing, Lu Yao, Zhang Ye, and Li Chun.

Diffusion-based generative models have long relied on Gaussian priors, with limited exploration of alternative distributions. The team introduced a Proper H?lder-Kullback Dirichlet framework, which uses time-varying multiplication transformations to define the forward and reverse diffusion processes. Unlike traditional reweighted evidence lower bounds (ELBO) or Kullback-Leibler upper bounds (KLUB), we propose two novel divergence metrics: Proper H?lder Divergence (PHD) and Proper H?lder-Kullback (PHK) divergence, with the latter aiming to restore the symmetry missing in existing formulas. When optimizing our diffusion model with PHK, we achieved a Fréchet Inception Distance (FID) of 2.78 on the unconditional CIFAR-10 dataset. Comprehensive experiments on image datasets validated the model's generative advantages and confirmed the effectiveness of PHK in model training. This work expands the diffusion model series with non-Gaussian processes and efficient optimization tools, offering new approaches for diverse, high-fidelity generative modeling.

Figure 3 Comparison of the samples generated by the Beta diffusion model and the Dirichlet diffusion model trained on the unconditional CIFAR-10 image dataset.

Paper link:

https://neurips.cc/virtual/2025/poster/116219

https://neurips.cc/virtual/2025/poster/118993