||Sourasekhar Banerjee <email@example.com>|
||2022-05-09 – 2022-12-01|
Federated learning adds a new dimension to the distributed machine learning by learning
models at the edge. The rapid emergence of edge devices, such as smartphones, tablets, wearable, etc., produces lots of complex data (i.e., non-IID, heterogeneous, etc.). Sending those complex data over the network is not communication efficient and also violates privacy constraints. Moreover, the edge devices are heterogeneous, so individual devices may not be suitable to learn a good model. So, federated learning could be helpful here. This project aims to research the learning models in federated settings on complex data and propose algorithms to optimize learning. The objective is two-fold 1) Multi-task federated deep learning, 2) Optimization of the federated algorithm for complex data. In objective one, we aim to design a feature selection algorithm for identifying useful features from data distributed across devices. The first objective
is to alleviate the statistical challenges by doing multi-task learning (MTL) in a federated
environment. Here, the goal is to solve an MTL problem when data is distributed across
devices. The second objective is to optimize the federated learning algorithms in non-convex settings to tackle statistical heterogeneity. The project also aims to give practical federated learning-based solutions for anomaly detection, image classification, etc.
The project needs hundred of Deep learning models to learn from multiple dataset (text/image). So it needs powerful GPUs to complete quickly.