The Federated Automated Deep Learning (FedAutoMoDL) project is a joint collaboration between Inria and IIIT-Delhi funded within the framework of the Inria Associate Team program.
Principle Investigators: Malcolm Egan (Inria) and Bapi Chatterjee (IIIT-Delhi).
While DNNs have had a huge impact on algorithm design for image and signal processing, it is well-known that the results are heavily dependent on the selected architecture. Typically, the architecture is tailored to a particular application by hand, which is susceptible to significant performance losses. Moreover, DNN architecture design has mostly been investigated in the context of image processing and natural language processing. For generic time-series data beyond speech and music signals (e.g., most sensor data), DNN architecture design remains in its infancy.
In federated settings, where data is not centralized due to privacy or communication constraints, high-performance design of DNN architectures is even less well-understood. This is now becoming a critical issue due to the widespread development of the IoT and edge computing.
The main goal of FedAutoMoDL is to develop algorithms for the systematic design of federated DNN architectures with a focus on scenarios with time-series data, such as that which often arises in the IoT and ranges from traffic management to data analytics. This will be achieved using tools from the emerging area of neural architecture search (NAS), and in particular the recently developed approach of differentiable neural architecture search.
