Zitan Chen

1 previous affiliation
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Zitan Chen focuses on distributed machine learning, specifically addressing data heterogeneity and generalization challenges in federated learning systems. Recent collaborative work investigates curriculum-based training strategies and prototypical calibration methods to improve model robustness across non-IID client distributions. This research emphasizes techniques for aligning feature spaces and managing sample complexity during distributed training.

Research Trajectory

In 2024, Chen's published work centered on overcoming the challenges of data heterogeneity in federated learning environments. A primary focus of this collaborative research has been the development of techniques to align feature spaces across distributed clients without compromising data privacy. For example, Chen contributed to Cross-Silo Prototypical Calibration for Federated Learning with Non-IID Data, which introduced a method called FedCSPC. This approach utilizes prototype information from clients to construct a unified feature space, mitigating the biases introduced by non-IID (independent and identically distributed) data and missing classes. This methodology intersects with the broader area of Latent Space Regularization for Class Separability, as it seeks to enforce structural consistency in the representations learned by isolated local models.

Building on the utility of class prototypes in distributed systems, Chen also collaborated on Cross-Training with Prototypical Distillation for improving the generalization of Federated Learning. This research addressed the problem of catastrophic forgetting, which often occurs when local models are cross-trained on heterogeneous data sources. The team proposed a prototype-guided cross-training mechanism (PGCT) designed to regularize changes in class-level data representations across clients. By stabilizing these representations, the work touches upon challenges found in the research direction of Resolving Semantic Confusion in Visually Similar Categories, ensuring that models maintain clear decision boundaries even when exposed to shifting local data distributions.

Alongside representation regularization, Chen's collaborative efforts have explored data-centric optimization strategies for federated systems. In Clustering-based curriculum construction for sample-balanced federated learning, Chen was part of a research team that investigated how the complexity of training instances impacts both local and global model performance. The resulting algorithm, FedAC, employs clustering-based curriculum construction to rank instance complexity and assign dynamic weights to training samples. This approach aligns with the theme of Sequencing Complexity for Robust Early Stage Training, demonstrating that carefully ordering the difficulty of examples can prevent local models from overfitting to heterogeneous noise, thereby yielding a more robust global model.

Current Directions

The most recent papers in the evidence, all published in 2024, indicate a sustained focus on the algorithmic foundations of federated learning. The central open problems addressed in this work revolve around reconciling the competing demands of local data privacy and global model generalization. By investigating both feature-space interventions—such as prototypical calibration and distillation—and data-centric approaches like curriculum learning, this recent work seeks to create more resilient distributed training paradigms capable of handling severe data heterogeneity across client silos.

Profile updated

See something wrong? Report a mistake