Most people who try to learn how convolutional neural networks work run into the same wall that user *cherry_190* hit. They can follow the math of forward propagation, they can write the backprop loop, but they cannot *see* the training. That wall is not a personal failing, it is a gap in how these concepts are taught. We believe that if a learner cannot visualize the process, the explanation is incomplete. Andrew Ng's course is excellent on structure. It is less excellent on showing you what "learning" actually looks like inside a model.
Here is what is happening during training that the slides often miss. Before training begins, a randomly initialized CNN treats every image as a meaningless grid of numbers. Its first guesses are pure noise, it will classify a cat and a car with equal, terrible confidence. Then you feed it a labeled image, compute the error, and nudge every filter and weight by a tiny amount so that the next guess is slightly less wrong. Repeat that across thousands of images, and the filters stop being random blobs. They begin to detect edges, then textures, then shapes, then parts of objects. The "same process" you see in every image run is the forward pass; the change you cannot see is each backward pass updating the filters by fractions of a percent. That accumulation of tiny corrections *is* the training. The model does not learn by being told about cats. It learns by being wrong about cats ten thousand times.
This matters for practical work because it reframes what debugging a model looks like. When a network fails, the temptation is to blame the architecture or the hyperparameters. More often the problem is that the model has not been shown enough examples of the *wrong thing* to recognize the right one. The visual that helped us was to think of the training process as a sculptor looking at a block of marble through a changing set of lenses. Each epoch adjusts the lens, and each batch gives the sculptor a slightly better view of where to chip away. Training loss does not descend because the model is "understanding" data. It descends because the model is iteratively removing its own bad guesses.
If you are trying to learn this material and feel stuck on visualization, step away from the notebook and do something concrete. Take a tiny 8x8 grayscale image, initialize three random filters, and walk through one update by hand on paper. You will see the filter values change. You will watch them become less random. That is the whole mechanism. The rest is scale.