The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networks

Neural Information Processing Systems 

Do neural networks, trained on well-understood algorithmic tasks, reliably rediscover known algorithms for solving those tasks?