modeling_basics.md

Modeling: Good default choices

For all problems

• rmsprop as optimizer
• Dense layers
• relu activations for intermediate layers

If data is limited

• k-fold cross-validation can help with better estimates of out of sample performance

By task

task	binary	multiclass	regression
loss	binary_crossentropy	categorical_crossentropy	mse
metrics	accuracy	accuracy	mae
output	1, sigmoid	k, softmax	1, no activation