The Theoretical ML team has outlined the following line of research:
- Make sure that everyone reaches (by studying the selected material) a solid shared theoretical knowledge about neural networks or other architectures the team plans to employ;
- Identify an appropriate dataset: it should be sufficiently large and with the possibility of being simulated or approximated well enough with a mathematical model;
- Pinpoint the architectures to concentrate on and test them on the selected data. Investigate their accuracy and generalization theoretical limits through stress tests, and besides, exotic phenomena from which it is possible to deduce theoretical properties which are much harder to find via pure mathematical studies;
- The goal is to publish a paper, in case the team finds significant results. We haven’t found any competition that suits this kind of research yet, but the possibility of taking part in contests in the future is not prevented.