Compiling...
So far, we have explored Feed-Forward Networks (great for spreadsheets) and Convolutional Networks (great for images). Both of these architectures share a massive limitation: They have absolutely no memory.
If you feed an image into a CNN, it processes it and forgets it immediately. If you feed the exact same image into the CNN again, it processes it from scratch. It treats every single input as completely independent.
But what if your data relies on the past?
