Introduction
The LSTM works brilliantly but has a lot of moving parts: two state vectors and four sets of weights. In 2014, Kyunghyun Cho and colleagues proposed the gated recurrent unit (GRU), a streamlined version that keeps the key idea (gates that control memory) with fewer pieces.
Think of a whiteboard instead of a notebook. With a notebook (LSTM), you have separate decisions for crossing things out and for writing new things, plus a separate decision about what to read out loud. With a whiteboard (GRU), there's a single decision: "how much of this spot do I erase and replace with the new information?" Whatever you don't erase stays. And what's on the board is directly what you show everyone.