Introduction
Recurrent cells process a sequence one step at a time: step 50 must wait for step 49. That makes them slow to train on modern hardware, which is built for doing many things in parallel. And even LSTMs and GRUs struggle with sequences of thousands of steps.
Convolutions offer another way. In Computer Vision you slid a small filter across an image in two directions. For sequences, you slide a filter along one direction: time. Every position can be computed at the same time, and by stacking layers cleverly, the network can see very far back.
Think of scanning a long receipt with a small window that shows 3 lines at a time. One pass finds patterns within each group of 3 lines. A second pass, over the results of the first, effectively sees 5 lines; a third sees 7. If each pass skips lines (looking at every 2nd, then every 4th line), the window's reach grows exponentially with each pass.