NVIDIA Deep Learning Institute: Fundamentals of Deep Learning Practice Test

An NLP engineer trains a unidirectional LSTM language model on 50-token sequences. After convergence, the team adds a second LSTM layer on top of the first (stacked) and also wraps both with dropout=0.25 between time steps and between layers. They observe that the validation perplexity improves by 4 points relative to the single-layer baseline, but training perplexity is now higher than validation perplexity for the first 15 epochs. What is the most accurate interpretation?

An ML engineer is implementing a CNN for 32x32 RGB CIFAR-10 images. The first layer is a 3x3 convolution with 32 filters, stride=1, padding=1, followed by ReLU and 2x2 max-pooling with stride=2. When the input resolution is doubled to 64x64 while keeping every other layer identical, what is the spatial size of the feature map immediately after the first pooling layer?

A researcher trains a two-hidden-layer MLP on the Iris dataset using stochastic gradient descent with a learning rate of 0.01. The model reaches 98% training accuracy but the decision boundaries appear overly smooth and fail to capture the non-linear separation between two of the classes in a held-out visualization. Which modification to the network architecture would most directly increase its capacity to model more complex decision boundaries?

You are preparing the final fruit project submission. The assessment script evaluates your model on a hidden test set of 200 fresh/rotten images that have never been seen by any student. Your local validation accuracy (with your own 80/20 split + augmentation) is 93%. On the hidden test the accuracy drops to 71%. Which practice from the DLI final-project guidance would have most reliably predicted this drop before submission?

What does the "batch" in batch normalization refer to during the training forward pass, and why must the implementation behave differently at inference time?