Ruinan Jin

Ruinan Jin

Machine Learning Research Student

About Me

Hello, I’m Ruinan Jin, and I am currently the PhD student focus on medical and trustworthy machine learning. I hold a B.S. (Hon.) in Computer Science from the University of British Columbia (UBC). Below is a list of my research interests, and I am open to various forms of collaboration. If you share any of these interests, please feel free to reach out! Thank you :)

Research Directions and Past Experiences

Federated Learning

  • Data Heterogeneity: personalized and heterogeneous federated learning strategies.
  • Distributed Generative Models: integration of generative adversarial networks (e.g., StyleGAN2-ADA) into federated settings.
  • Privacy: differential privacy techniques under federated learning.
  • Data Distillation: data distillation approaches, such as gradient matching and distribution matching, under federated learning.
  • Federated Recommendation System: integrate deep learning recommendation model (DLRM) into heterogeneous federated learning for efficient communication.
  • New Regime of Federated Learning: incorporation of textural gradient features into federated optimization.

Trustworthy Machine Learning

  • Security: membership-inference attacks, patch / Fourier backdoor insertion, robust aggregation, and certified defense frameworks.
  • Privacy: central- and local-model differential privacy plus private synthetic-data generation for sensitive domains.
  • Memorization and Unlearning: memorization behavior of modern vision-language models and its broad implication in the trustworthy machine learning (e.g., backdoor attack, fairness, out-of-domain data and privacy), along with the development and application of machine unlearning algorithms.
  • Fairness: bias quantification and group-fairness constraints in medical foundation models and vision-language regimes.

Vision-Language Multimodal Models

  • Contrastive Pre-training: CLIP-family for image-text representation alignment.
  • Generative VL Models: LLaVA-family for visual question answering, captioning, and multimodal reasoning.
  • Generalization: in domain and out-of-domain performance benchmark in medical vision-language models.
  • Compositional Learning: usage of the textual concepts (e.g., concept bottleneck model) to predict intermediate human-interpretable attributes before final decisions.

Software

FairMedFM: Collection of Foundation Models for medical image analysis and their fairness usages.