Overview
This is the MVP I built for our federated deep learning healthcare AI startup back in 2019. We demoed this MVP to several innovation leads and heads of data science at pharma companies. This MVP contained the minimum feature set of a federated deep learning platform. It allowed data scientists to:
- Connect to a hub node running Jupyter (master EC2 node).
- From the hub, connect to three independent hospital servers (modelled as separate EC2 instances).
- Query the predefined intensive care unit datasets of these hospitals.
- Specify a dataset and corresponding target variable to train and test on.
- Specify a deep learning model architecture in PyTorch.
- Select a cost function to optimize and performance metrics to track.
- Specify various training parameters such learning rate and number of local training steps.
- Perform federated stochastic gradient descend with weight averaging to learn a global model
Further details
Please make sure to check out my blog post about the story of our startup and how / why we failed.