Who Does the Work: Designed Pipeline vs. Learned Pipeline | Visual AI Learning Guide

Look at the two pipelines side by side. Both start with the same raw input on the left and end with a prediction on the right, so the ends match. The difference is in the middle. In the top pipeline, a person sits between the raw data and the model: they decide which properties of the input matter and hand those chosen features to the model. The human judgment is applied before the model ever sees the data. In the bottom pipeline, the raw input goes straight to the model, and the model itself works out an internal representation of what matters before it predicts. Here the human judgment is applied later, in how the model is built and trained rather than in what it is allowed to look at. So the split is not about whether a human is involved, but about where in the process their judgment enters.

0:00 / 0:00

Auto-play

The central philosophical difference between the two approaches is where the human effort sits. In a traditional machine learning pipeline, a person inspects the raw data, decides which properties are relevant, and encodes them as features. The model then learns a mapping from those human-chosen features to the output. The human decides what the model is allowed to look at.

In a deep learning pipeline, the raw input is fed to the model, and the model learns its own internal representation of what matters before producing the output. The human still makes decisions, but they concern the data, the architecture, and the training process rather than the specific properties of the input. The human decides how the model learns, not what it looks at.

Both pipelines still end in a prediction, and both still require labeled examples. The difference is not whether a human is involved, but at which stage the human's judgment is applied: up front, in designing the input representation, or later, in shaping how the model learns its own.