The Skeleton in the Gaps
In the dusty museum workroom, a junior curator stared at the large table. It should have been covered in fossils, but only a few leg bones and vertebrae sat there. About three-quarters of the dinosaur was missing. This wasn't a mistake, though. It was a deliberate test set by the head paleontologist.
Usually, apprentices learn on skeletons that are nearly complete. If a toe is missing, they just copy the one next to it. It's easy, but it makes them lazy. They stop thinking about how the animal actually moved and just match patterns, like patching a crack in the pavement without checking the foundations.
The boss had a different plan. She swept away most of the bones, leaving only a quarter behind. She told the student that to rebuild the beast from scraps, they couldn't rely on nearby clues. They had to understand the whole creature's structure. A single hip bone had to tell them exactly how long the leg should be.
This changed the work completely. The student only had to clean and examine the few bones actually on the table, ignoring the empty space. Since they weren't wasting time on the missing parts yet, the first check went three times faster. They could get through far more skeletons in a single afternoon.
Once the real shards were analysed, the student switched to sculpting. They had to predict the missing seventy-five per cent from scratch. Because the task was so hard, the final result proved they had mastered the 'grammar' of the skeleton far better than if they had just connected the dots on an easy job.
It turns out computers learn the same way. When we force them to imagine a whole picture from just a few hints, they stop memorising surface details. By hiding most of the information, we teach machines to understand the true essence of a scene, just like that paleontologist teaching her student to see the dinosaur.