Hello everyone!
I’d like to discuss 1st place solution of Planet: Understanding the Amazon from Space Kaggle competition. So far I’ve taken only the first part of the course and so I find it a bit challenging understanding the approach taken by the author.
I understand the preprocessing part and the use of multiple CNNs, but the ensembling step is unclear.
In particular, I do now understand the use of two ridge regressions. What is the purpose of the ridge regression and what is intuitive understanding of what it is doing in this case?
The author says:
I then passed each CNN’s class label probabilities through its own ridge regression model, in order to adjust the probabilities to take advantage of label correlations.
How does ridge regression take advantage of that?
Then:
Finally, I ensembled all 11 CNNs, by using another ridge regression model.
What is the purpose of this step?
I would be happy to hear your feedback and in general any other thoughts about this competition.