COVID-19 has changed our lives in many aspects. Not only our social life was upended, but our work priorities as well. Our team was working to create the next generation of community machine learning when the pandemic hit. So, instead of waiting for the storm to end, we decided to make an early release and join the fight against COVID-19.
Since we are building a competitive AI crowdsourcing platform, the path was clear: we should use our resources to host a community-made algorithm, helping those who are dealing with the virus every day. We have teamed up with researchers from FIMM (Institute for Molecular Medicine Finland, HiLIFE) and the Biological Research Centre of Szeged, Hungary to create a novel serology-based test for diagnosing the disease in several stages.
Here is how you can participate!
Building a COVID-19 diagnostic tool
Diagnosing COVID-19 in a fast and reliable way is difficult. Rapid tests can be inaccurate, while accurate tests can take a while and require complicated experimental protocol. So, there is much to improve.
A fast and accurate test can be performed with blood tests. Simply put, blood is drawn from a potential patient, which is used to grow a cell culture in about an hour. Depending on some antibodies present, we can tell if the patient has active disease, cured of the disease, or have not been infected yet.
The diagnosis is done with microscopy. It is the real challenge: predicting the presence of COVID-19 from the cellular images itself. This is where you come in. Your task is to train a classifier, which is capable to recognize which sample is infected.
To enter the competition, you should submit the results of your model on the test set. We score your submission and place it on the leaderboard. Each week, the best model will be deployed to production, available for everyone to use.
If you are interested, sign up for the competition at https://telesto.ai!
A new and dynamic competition format
Why are we deploying the best model to production every week?
In machine learning, development is not a linear process, but rather a cycle. After a model has been delivered, it is continuously tested in production. When issues arise, the dataset is enhanced and a new model is trained.
The current crowdsourcing model is not a good fit for this. So, we have designed a dynamic competition format to solve these problems. Instead of having a few winners in the end, the top solutions are rewarded and deployed each week. This way, the entire development cycle becomes part of the crowdsourcing.
Our COVID-19 competition is a public test of the idea. Join the competition and help us develop the next generation of community machine learning!