I wrote myself a simple commitment this month: learn machine learning, and use the channel to keep track of the process. The difficult part is turning that broad intention into a sequence of questions small enough to answer.
Coming from graphics, I have a useful place to begin. An image is an output I can inspect. Coordinates, sampling, color spaces and transformations are already concrete ideas. A neural network introduces a different way to construct part of that pipeline, but it doesn't remove the need to understand the data entering and leaving it.
My first goal would be to understand one supervised problem from end to end. For example, generate a clean image, add a known kind of noise, and train a small model to predict the clean version. The target is unambiguous because the data-generation process supplies it. That makes the exercise more manageable than asking a network to improve arbitrary pictures according to an undefined idea of quality.
The baseline should be deliberately ordinary. A simple blur can remove some noise too. If the learned model looks better, I want to know where: on edges, in flat regions, or in textured detail. If it fails, a difference image may explain more than a single overall score.
There is a familiar graphics trap here. An attractive result can distract from the wrong transformation. A model that smooths away small features may produce a cleaner-looking picture while discarding information. A sharpening process can make an image feel more detailed without recovering the true detail. I want to keep perceived quality and reconstruction accuracy as separate questions.
The work on image-to-image translation is one reason the subject feels relevant to graphics. It treats the relationship between two visual representations as something a model can learn. But it also makes the dataset and objective impossible to ignore. Edges to photographs, noisy to clean, and one rendering style to another are different tasks even when the network diagrams look similar.
For learning, I would keep the train and test sets separate from the beginning. Reproducing a familiar training image is not the same achievement as handling a new example. With rendered data, it is especially easy to create near-duplicates by changing a camera slightly and accidentally place almost the same scene in both sets.
A useful notebook should therefore record more than the final image: what changed, what remained fixed, which examples failed and what comparison would challenge the explanation. That is the habit I want to develop before making the models larger.
There are plenty of impressive demos to collect. My next step is smaller: understand one model well enough to predict how a change in its inputs, loss or training data should affect the output. If I can explain an ugly result, I will probably learn more from it than from another beautiful screenshot.