Every month, scientists measure ocean surface temperatures and produce a map. In the tropical Pacific, a warm patch sometimes appears: El Niño, which alters weather worldwide. Teaching a computer to find this patch on a map proved tricky. Conventional neural networks require lots of data and long training times. But there is a special type of network, the echo state network, which runs fast and needs almost no tuning. The problem was that no one could explain how it reaches its decisions.

You can picture an echo state network as a pond into which you drop pebbles. Each pebble is a column of the temperature map, and the ripples on the water are the network's response. Training only touches the output layer, so there are few parameters: 301 instead of nearly 88,000 for a conventional network. The authors applied a method previously used for deep networks: they traced backward from the answer and saw which pebbles caused the strongest ripples.

The scientists ran a simple experiment: they shuffled the columns of the map randomly. The network still found the right region, even though the spatial structure had vanished. This means the echo state network does not rely on pixel positions but extracts features from any data order. However, the reservoir has a parameter that determines how long it remembers the past: the leaking rate. When they increased it from 0.01 to 0.4, accuracy dropped to 58%, and the relevance maps showed that the network had forgotten early columns and looked only at the end of the map. Now echo state networks are no longer a black box, and climatologists can trust their conclusions.