3 ms·
Can you please comment on why you need a 50 node 3 hidden layer ffnn to do regression, as opposed to something simpler ?
by dxbydt 12y ago
Can you please comment on why you need a 50 node 3 hidden layer ffnn to do regression, as opposed to something simpler ?
- larrydag 12y agoOne reason could be that its a non-linear prediction.
- minimaxir 12y agoYou can do simple regressions (linear, logistic) natively and easily in R. The tutorial is demonstrating H2O by using an algorithm (deep learning) that's not native to R.
- jofai_chow 12y agoThe starter code I provided is a basic DNN structure for modelling complex non-linear relationships between five soil properties and 3000+ predictors. In practice, I found that some of the properties require even more complex DNN structure to achieve better predictive accuracy. The 50-50-50 setup is a very solid starting point for the readers to begin their own experiments.
- dxbydt 12y agoThank you. How did you come up with the 50-50-50 setup, or was it purely empirical ? Did you try something simpler first, and how did that simpler method perform vis-a-vis this DNN ? Congratulations on topping the leaderboard.
- jofai_chow 12y agoThanks! Yes, I always start with much simpler networks like 10, 10-10, 10-10-10. Unfortunately, the regression problems here are quite complex hence bigger networks are required (well, it wouldn't be on Kaggle otherwise).