Echo State Network

A recurrent neural network architecture that simplifies training by using a fixed, random reservoir to map temporal inputs into high-dimensional feature spaces.

Echo State Networks (ESNs) solve the vanishing gradient problem in recurrent models by freezing the internal weights of a large, sparsely connected reservoir. Only the linear output layer is trained, typically via ridge regression, which makes the learning process computationally cheap and exceptionally fast. Developed by Herbert Jaeger in 2001, ESNs excel at chaotic time-series prediction (like the Lorenz system) and wireless channel equalization. They rely on the Echo State Property: the effect of previous inputs must gradually vanish so the current state is a unique function of the input history. This makes them a go-to choice for real-time hardware implementations and low-power edge computing.

https://scholarpedia.org/article/Echo_state_network