7. Examples and Applications
A ResNet adds spatial layouts and skip connections to the operations used by the MLP. GPT adds token streams and causal masks. Fourier neural operators connect learned maps to PDE data, while diffusion and reinforcement learning introduce probabilistic transitions and evolving state. We can still trace a prediction through the familiar layer operations, provided we also account for these additional choices.
The extra structure changes what we need to inspect. For attention, two tensors can have the expected dimensions while using the wrong causal mask. For a Fourier layer, a transform convention or selected mode range affects the represented operator. A diffusion step depends on a time index and noise schedule as well as learned weights. The examples keep these choices close to the code and outputs so that the familiar layer operations remain connected to the actual model being run.