[HugThem]

    statistics is about explaining an observed
    outcome in terms of its causing factors.

Isn't it the same for ML? When we train a NN to map input factors to the observed outcome, it will set the weights of the factors that do not influence the outcome to 0.

[thom]

This is only the case if your model’s features are easy to understand, and one of the motivations for some NN approaches is to avoid having to create your own features.