Neural Networks and Random Forests

Retraining neural networks can recover diversity but at substantially higher wall-clock cost than GP conditioning. Random forest degeneracy shows that merely incorporating new data is insufficient for pseudo-observation batch selection. Re…

1 sources - 4 claims

Retraining neural networks can recover diversity but at substantially higher wall-clock cost than GP conditioning. Random forest degeneracy shows that merely incorporating new data is insufficient for pseudo-observation batch selection. Rebuilt random forests remain degenerate because bootstrap sampling dilutes the effect of one pseudo-observation. Non-GP alternatives such as neural networks can cause batch points to collapse to identical locations under pseudo-observation selection.