PROJECT TITLE :

Dynamic Neural Graphs Based Federated Reptile for Semi-supervised Multi-Tasking in Healthcare Applications

ABSTRACT:

AI healthcare applications are dependent on confidential electronic health records (EHRs), which are often dispersed across a network of symbiont institutions and have little to no labeling. It is difficult to train effective machine learning models on such data because of the challenges involved. As a solution to these problems, the work presented here proposes a federated learning framework that is based on dynamic neural graphs. A model agnostic meta-learning (MAML) algorithm known as Reptile is extended to work in a federated environment by the framework that has been proposed. This paper, however, proposes a dynamic variant of neural graph learning (NGL) to incorporate unlabeled examples in a supervised training setup. This is in contrast to the MAML algorithms that are already in existence. A meta-learning update is computed by Dynamic NGL by carrying out supervised learning on a labelled training example while simultaneously carrying out metric learning on the labelled or unlabelled neighborhood of the training example. This neighborhood of an example that has been labelled is determined in a dynamic manner by using local graphs that are constructed over the batches of training examples. The construction of each local graph involves determining the degree to which different embeddings produced by the model's current state are similar to one another. The incorporation of metric learning on the neighborhood transforms the nature of this framework into one that is only semi-supervised. The experimental results obtained using the MIMIC-III dataset, which is accessible to the general public, demonstrate how effective the proposed framework is for both single-task and multi-task environments, even when data decentralization is restricted and limited supervision is available.