Program Overview
This research program aims to develop a robust federated learning (FL) framework tailored for seismic interpretation tasks, particularly in the presence of data heterogeneity and limited label availability. Seismic data often comes from geographically dispersed regions or diverse partitions of the same volume, making it an ideal candidate for decentralized training. However, such variation in data distribution across clients poses critical challenges, including reduced model accuracy, increased forgetting of prior knowledge, and poor personalization of the global model to local clients.
Building upon prior work from the OLIVES lab in the domain of catastrophic forgetting in medical datasets, this program proposes to extend and adapt those insights to seismic data. Specifically, the research focuses on mitigating forgetting at the client level to improve both personalization and generalization in federated models. Furthermore, the program will develop new aggregation strategies that prioritize clients with higher generalization capabilities, fostering better-balanced global models.
Objectives
- Quantify the effects of data heterogeneity on forgetting and model degradation in federated learning setups.
- Develop client-side training strategies to mitigate forgetting and enhance personalization.
- Design novel aggregation mechanisms that adaptively weight client contributions to improve generalization.
- Validate the FL framework on seismic (and medical datasets) to ensure cross-domain applicability.
Key themes
- Data Heterogeneity and Forgetting: Understanding how distributional shifts affect model memory and performance.
- Personalization in Federated Learning: Enhancing client-specific adaptation without sacrificing global coherence.
- Generalization through Adaptive Aggregation: Improving global model robustness via selective client weighting.
- Cross-Domain Validation: Ensuring that methods generalize across seismic and medical imaging applications.
Participation and Governance
This CRP operates within the ML4Seismic philanthropic partnership framework. Participation requires enrollment as an Executive Member with at least one CRP selected. Partners are
invited to:
- Participate in CRP-specific workshops and benchmarking initiatives
- Collaborate on shared scientific objectives and data challenges
- Engage with students and researchers in community events
All research outcomes are shared via open-source repositories and peer-reviewed publications in accordance with ML4Seismic operating guidelines.
Contact
Prof. Ghassan AlRegib alregib@gatech.edu
