Research
My current project is about multi-model federated learning (MMFL): local clients can handle multiple unrelated FL models concurrently.
For example, a smartphone user can generate data for both keyboard prediction and speech recognition models.
Instead of building two single-model FL systems, we build one MMFL system and solve system-level optimization problems to make it more efficient.
|
My research at CMU: Federated Learning and Optimization
|
[On-going project] Client sampling in a more heterogeneous MMFL scenario. How to optimize the sampling distribution?
|
Group-based Client Sampling in Multi-Model Federated Learning
Zejun Gong*,
Haoran Zhang*,
Marie Siew,
Carlee Joe-Wong,
Rachid El-Azouzi
Under Review (ICASSP), 2024. *Equal contribution
|
Poster: Optimal Variance-Reduced Client Sampling for Multiple Models Federated Learning
Haoran Zhang,
Zekai Li,
Zejun Gong,
Marie Siew,
Carlee Joe-Wong,
Rachid El-Azouzi
ICDCS, 2024 [Best Poster Award]
[Poster]
Supplementary material
This paper explores an MMFL scenario where each client is restricted to training only one model per global round due to limited computational resources.
We propose a client sampling and task assignment method that theoretically minimizes the variance of global updates. The method achieves a 30% higher accuracy than the baseline.
|
Fair Concurrent Training of Multiple Models in Federated Learning
Marie Siew,
Haoran Zhang,
Jong-Ik Park,
Yuezhou Liu,
Yichen Ruan,
Lili Su,
Stratis Ioannidis,
Edmund Yeh,
Carlee Joe-Wong
arXiv, 2024
Supplementary material
In the real-world scenario, clients could contribute to multiple federated learning models' training at the same time.
For example, a company may update their keyboard prediction model, speech recognition model, and other more federated learning models on your phone.
In this paper, we propose a client-task assignment strategy to preserve the fairness across multiple training tasks.
|
My research at HKUST: Medical AI
|
Efficient 3D Transformer with cluster-based Domain-Adversarial Learning for 3D Medical Image Segmentation
Haoran Zhang, Hao Chen
ISBI (one of the best conferences in medical image processing), 2023
We proposed a pyramidally downsampled 3D Transformer to boost efficiency and maintain high segmentation accuracy.
This paper also discussed the domain generalization problem, and
proposed a cluster-based domain-adversarial learning method to exploit available domains at a fine-grained level.
|
About Me
I was born in Xinxiang, China.
My family name is 张 (Zhang), which is one of the most common family names in China.
And my given name is 皓然 (Haoran), 皓(Hao) means "the color of the moon", 然(Ran) doesn't have a specific meaning, maybe it means "yeah, uh-huh".
As the first in my family to attend college, I am deeply grateful for the support from my parents, mentors, and friends.
|
|