Fig. 5

Figure shows the impact of the choice of penalty on model performance. The bar plots indicate the average cross-validation balanced accuracy in multiple runs. In all experiments a regularization of the form r ₁−r ₂ has been applied where r ₁ indicated the regularization applied to the weights in the first layer W ⁽¹⁾ and r ₂ indicates the regularization applied to the weights in the second layer W ⁽²⁾. Half L2: ℓ ₂-Null), Half L1: Null- ℓ ₁, Full L1: ℓ ₁- ℓ ₁, Full L2: ℓ ₂- ℓ ₂ and GRRANN: ℓ ₂- ℓ ₁