CAEFL: composable and environment aware federated learning models

Abstract

Federated Learning allows multiple distributed agents to contribute to a global machine learning model. Each agent trains locally and contributes to a global model by sending gradients to a central parameter server. The approach has some limitations: 1) some events may only occur in the local environment, so a global model may not perform as well as a specialized model; 2) changes in the local environment may require an agent to use some dedicated model, that is not available in a single global model; 3) a single global model approach is unable to derive new models from dealing with complex environments. This paper proposes a novel federated learning approach, CAEFL, that is local environment aware and composes new dedicated models for new complex environments. CAEFL is implemented in Elixir to exploit transparent distribution, pattern matching, and hot-code-swapping. Pattern matching is used to transform environment sensors data to corresponding tags and aggregate data with the same environment tags on agents. It is also used on parameter server to match client’s push/pull request for these tagged models. It enables a declarative way for environment aware federated learning approach. CAEFL outperforms state of the art federated learning by 7-10% for the MNIST dataset and 2% for the FashionMNIST dataset in specific and complex environments.