This repository contains my final submission to the AutoML decathlon competition which won 4th place. Access the competition page here.
This flowchart illustrates the model selection strategy:
Main rules I tried to follow:
- Don’t overfit to the train tasks
- Keep models simple.
- Try not to look at the data.
- Reduce the number of moving parts / isolate improvements.
- Be wary of time and memory constraints
- Check how much time you have left and stop training accordingly.
- There is no free lunch
- Inductive bias is important. Pick the right model class for every type of task.