A model trained on ~10,000 controlled images fits the training set well but performs poorly on unseen test images (training error decreases with epochs while test error remains high). Which actions will help address this problem? (Choose two.)

The problem described, where training error decreases but test error remains high, indicates overfitting. The model has learned the training data too well, including noise, and fails to generalize to new, unseen data. Data augmentation increases the diversity of the training set by creating modified versions of existing images (e.g., rotations, flips, crops). This exposes the model to more variations, making it more robust and less likely to overfit to specific training examples. L2 (weight) regularization adds a penalty to the loss function based on the magnitude of the model's weights. This encourages the model to use smaller weights, which results in a simpler model that is less prone to overfitting. Adding vanishing gradient is not a solution; vanishing gradients are a problem to be solved, not a technique to apply. Increasing network complexity would exacerbate overfitting. Gradient checking is a debugging technique for verifying backpropagation, not a solution for overfitting.