Wei Ding

Deep neural networks are increasingly deployed in biomedical and healthcare applications, where robustness, adaptive knowledge reuse, and reliable prediction on rare or difficult cases are critical. However, biomedical datasets often evolve over time and contain rare patterns that can be easily forgotten during model updates. As a result, current DNN model building suffers from two serious problems: forgetting and difficult low-confidence test cases. In this paper, we propose an open-style skill acquisition approach that contrasts with conventional closed-style training, where recently learned features may overwrite previous ones to minimize overall loss in backpropagation. Instead of compressing all knowledge into a single fixed model, our approach accumulates diverse skills (represented as groups of neurons and connections) during training to reduce the forgetting problem. During inference, the model dynamically selects the most suitable skills for each test sample to maximize prediction confidence, thereby addressing difficult low-confidence cases during inference. We evaluate our approach on multiple benchmark datasets and biomedical image datasets from the MedMNIST benchmark. Experimental results demonstrate consistent performance improvements over state-of-the-art ensemble methods, highlighting the potential of open-style skill acquisition for robust biomedical deep learning.

Chaotic dynamics are ubiquitous in many real-world systems, ranging from biological and industrial processes to climate dynamics and the spread of viruses. These systems are characterized by high sensitivity to initial conditions, making it challenging to predict their future behavior confidently. In this study, we propose a novel deep-learning framework that addresses this challenge by directly exploiting the long-term compounding of local prediction errors during model training, aiming to extend the time horizon for reliable predictions of chaotic systems. Our approach observes the future trajectories of initial errors at a time horizon, modeling the evolution of the loss to that point through the use of two major components: (1) a recurrent architecture (Error Trajectory Tracing) designed to trace the trajectories of predictive errors through phase space, and (2) a training regime, Horizon Forcing, that pushes the model’s focus out to a predetermined time horizon. We validate our method on three classic chaotic systems and six real-world time series prediction tasks with chaotic characteristics. The results show that our approach outperforms the state-of-the-art methods.

Deep learning models built upon Transformer architectures have led to substantial advancements in sequential data analysis. Nevertheless, their direct application to video-based tasks, such as Group Activity Recognition (GAR), remains constrained by the quadratic computational complexity and excessive memory requirements of global self-attention, especially when handling long video sequences. To overcome these limitations, we propose SUGAR: A Sequence Unfolding Based Transformer Model for Group Activity Recognition. Our approach introduces a novel sequence unfolding and folding mechanism that partitions long video sequences into overlapping local windows, enabling the model to concentrate attention within compact temporal regions. This local attention design dramatically reduces computational cost and memory footprint while maintaining high recognition accuracy. Within the Bi-Causal framework, SUGAR replaces conventional Transformer blocks, and experimental results on the Volleyball dataset demonstrate that our model achieves state-of-the-art performance, consistently exceeding 93% accuracy, with significantly improved efficiency. In addition, we investigate Lightning Attention 2 as an alternative linear-complexity attention module, identifying practical challenges such as increased memory usage and unstable convergence. To ensure robustness and training stability, we incorporate a dedicated safety mechanism that mitigates these issues. In summary, SUGAR offers a scalable, resource-efficient solution for group activity analysis in videos and exhibits strong potential for broader applications involving lengthy sequential data in computer vision and bioinformatics.

Large Language Models (LLMs) have shown strong general capabilities in many applications. However, how to make them reliable tools for some specific tasks such as automated short answer grading (ASAG) remains a challenge. We present SteLLA (Structured Grading System Using LLMs with RAG) in which a) Retrieval Augmented Generation (RAG) approach is used to empower LLMs specifically on the ASAG task by extracting structured information from the highly relevant and reliable external knowledge based on the instructor-provided reference answer and rubric, b) an LLM performs a structured and question-answering-based evaluation of student answers to provide analytical grades and feedback. A real-world dataset that contains students' answers in an exam was collected from a college-level Biology course. Experiments show that our proposed system can achieve substantial agreement with the human grader while providing break-down grades and feedback on all the knowledge points examined in the problem. A qualitative and error analysis of the feedback generated by GPT4 shows that GPT4 is good at capturing facts while may be prone to inferring too much implication from the given text in the grading task which provides insights into the usage of LLMs in the ASAG system.

This paper provides an overview of the ACM SIGKDD 2024 AI4Science4AI special day. It includes information about the organizers, invited speakers, keynote speakers, the event agenda, and insights from related workshops. The AI4Science4AI special day aims to bring together experts in artificial intelligence (AI) and science to discuss the latest developments, challenges, and future directions.

Mode collapse has been a persisting challenge in generative adversarial networks (GANs), and it directly affects the applications of GAN in many domains. Existing works that attempt to solve this problem have some serious limitations: models using optimal transport (OT) strategies (e.g., Wasserstein distance) lead to vanishing or exploding gradients; increasing the number of generators can cause several generators focusing on the same mode; and approaches that modify the loss also do not satisfactorily resolve mode collapse. In this article, we reduce mode collapse by formulating it as a Monge problem of OT map. We show that the Monge problem can be transformed to the distribution transformation problem in GAN, and a rectified affine neural network can be considered as a measurable function. In this way, we propose Monge GAN that uses this measurable function to transform the generated data distribution into the original data distribution. We utilize the Kantorovich formulation to obtain the OT cost, which is regarded as the OT distance between the two distributions. Finally, we conduct extensive experiments on both image and numerical datasets to validate our Monge GAN in reducing model collapse.