A credit-risk model shows 99% training accuracy but only 82% validation accuracy. The company needs to fix this performance gap before deploying. Which measures address overfitting and improve generalization?

The large gap between training (99%) and validation (82%) accuracy indicates significant overfitting. Dropout layers randomly deactivate neurons during training, preventing complex co-adaptations and forcing the network to learn more robust features. L1 and L2 regularization add penalties to the loss function based on the magnitude of model weights, discouraging overly complex models. K-fold cross-validation helps assess model performance more reliably by training and validating on different subsets of the data, providing a more robust estimate of generalization error and helping to identify overfitting early. Adding more dense layers increases complexity, exacerbating overfitting. Using PCA or decreasing model layers might reduce complexity too much, leading to underfitting. Augmenting data or removing duplicates are good practices but don't directly address the model's complexity issue in the same way regularization techniques do.