Disperse.ai | Devpost

Inspiration

Training AI models for modern medicine entails stringent privacy requirements and rigorous oversight from medical professionals to be deployed in a real world setting. Costs for training traditional models and scientific teams are high and don't always result in accurate models and efficient timelines on top of spending time ensuring all privacy specs are met. With new approaches in ML, particularly Weavechain's Split learning stack, researchers can train and ship models across multiple nodes and not have to sacrifice privacy and model accuracy.

What it does

Disperse.ai harnesses Weavechain's Split Learning approach to split a large dataset into segments and train the data across multiple nodes all while the data remains private under a hash. This approach not only allows for a more decentralized approach for sourcing and training data but also enables for researchers to collaborate together on building out models while not sacrificing privacy.

How we built it

For this example, I took a data set of 33,000 images, split it into 3 groups, constructed a collection of 3 nodes to train the data, and then initiated the notes to train the model. However, to get the model to run, a smaller data set for Multiple Sclerosis was used. The backend stack was done using off the shelf tools like Docker, while the front end interface was run through Jupyter Notebooks.

Challenges we ran into

There were some hiccups early on in getting the nodes to communicate with one another. There was also a decent learning curve in understanding the concept of split learning. I've studied Swarm learning back when studying robotics and the concept was similar, however there was an added layer of complexity given that the data is published on chain and protected using hashes.

Accomplishments that we're proud of

Excited that the 1. the model was up and running and 2. the technique works smooth. Also proud to say that Weavechain's docs were smooth and the setup process stood up with nothing blowing up too much.

What we learned

Learned a lot about Split learning and decentralized training of models.

What's next for Disperse.ai

I'm a researcher for vector borne diseases and plan to apply this approach to a project I've been spearheading in the field of Malaria and tick-borne disease research. I think the next layer for this project would entail a wallet feature where patients can contribute their data and be compensated for each submission using a Web3 wallet like Weavechain, Coinbase of Safe.

WATCH VIDEO ON LOOM BC YOUTUBE WOULDNT UPLOAD https://www.loom.com/share/ef07217077704eacaab553dc5e6238cd

VIMEO LINK https://vimeo.com/846606865?share=copy Password: Weavechain!