Drawings in untrained neural networks survive the training process

Researchers found that drawings embedded in randomly initialized neural networks can persist through the training process, suggesting that certain patterns in untrained networks are robust enough to survive optimization.