What happened
Researchers discovered that what matters for training deep image-recognition neural networks isn't how many layers you stack, but how many actual computational steps the information has to travel through. This distinction explains why some architectural designs like ResNet can be made very deep without breaking, while others like VGG hit a wall — and it gives engineers a principled way to design better networks without guessing.
Why it matters
For years, engineers assumed 'deeper networks are better' because they have more layers, but they kept hitting diminishing returns and training failures on very deep models. This paper shows the real bottleneck isn't layer count — it's the path information takes through the network during training. If you can decouple those two things (which modern architectures do via shortcuts), you can scale much further. That's a conceptual shift that lets engineers stop designing by intuition and start designing by measurement.