One task, two routes. The task is deciding whether a photograph contains a cat. The question is not which route wins, but where the human effort goes and what each route can accept as input.
Where the work sits
Handcrafted features
- A person studies many photographs and names the distinguishing properties: ear shape, fur texture, eye arrangement.
- Each property is computed by hand-written code and passed to the model as a number.
- The model combines those numbers into a decision but never sees a pixel.
- A property left off the list is invisible to the model and cannot be recovered by training.
- The human's reasoning is explicit and can be inspected and explained.
Learned features
- Raw pixels are passed directly to the model.
- The layers build their own intermediate descriptions, from simple local patterns up to larger structures.
- No code is written to detect ears or fur; the model discovers what matters.
- The route demands large amounts of labeled examples and substantial compute.
- The internal descriptions are opaque to the person who built the model.
The two routes are not ranked; they relocate the human effort. Handcrafted features put the effort into specifying what matters and keep the reasoning inspectable. Learned features remove that specification work but require more data and compute, and give up transparency in return.
How much data each route needs, and which route is the better practical choice, are separate questions. This page only compares where the features come from and what that costs.