A health care company trains a neural network to classify X-rays as normal or abnormal. The dataset contains 1,000 labeled training images and 200 test images. A network with 50 hidden layers achieves 99% accuracy on the training set but only 55% accuracy on the test set. Which of the following changes should the specialist consider to address this problem? (Choose three.)

The large discrepancy between training accuracy (99%) and test accuracy (55%) indicates significant overfitting. Reducing the number of hidden layers simplifies the model, making it less prone to memorizing training data and improving generalization. Increasing hidden layers would exacerbate overfitting. Enabling dropout regularization randomly deactivates a percentage of neurons during training, preventing complex co-adaptations and forcing the network to learn more robust features. Enabling early stopping monitors performance on a validation set (or the test set, if a separate validation set isn't available, though a dedicated validation set is best practice) and halts training when performance starts to degrade, preventing the model from overfitting further. Using a smaller learning rate primarily affects convergence speed and stability, not directly addressing overfitting in this scenario. Including test images in the training set would invalidate the test set's purpose as an unbiased evaluation of generalization.