Pluralis Research focuses on foundational research in Protocol Learning, enabling multi-participant training of large foundation models without any single entity holding the complete model weights. This research facilitates the creation of community-trained and community-owned frontier models with sustainable economics. Key contributions include communication-efficient model parallelism and optimization techniques for decentralized training across low-bandwidth networks.
Funding
Funding not disclosed
Founders
Product
Problem
Centralized training of foundation models requires organizations to pool sensitive data, raising privacy and security concerns. Traditional approaches struggle to balance model performance with the need to protect proprietary or confidential information during the training process. This limits collaboration and innovation in developing AI models across distributed datasets.
Solution
Pluralis is developing a decentralized AI protocol that enables organizations to collaboratively train foundation models on distributed data sources without directly sharing the underlying data. This approach leverages techniques like federated learning and secure multi-party computation to facilitate model development while preserving data privacy and security. By decentralizing the training process, Pluralis aims to unlock the potential of distributed datasets and foster collaborative AI innovation without compromising data confidentiality. The protocol allows for efficient model development while adhering to stringent data protection requirements.
Target Audience
The primary target audience includes organizations and research institutions that require collaborative training of AI models on distributed datasets while maintaining data privacy and security.
Features
- Decentralized training protocol for foundation models
- Support for federated learning and secure multi-party computation
- Data privacy and security mechanisms to protect sensitive information
- Collaborative model development across distributed data sources