Reservoir computing is a shortcut for training recurrent neural networks. Instead of adjusting every connection with backpropagation through time, which is slow and can be unstable, you build a large recurrent network with random, fixed weights, called the reservoir, and drive it with the input signal. Each unit ends up carrying a different nonlinear echo of recent inputs. Then you train only the output layer, usually with plain linear regression, to combine those echoes into the output you want.
The idea arose twice, independently, around 2001–2002: Herbert Jaeger’s echo state networks, built from ordinary artificial neurons, and Wolfgang Maass’s liquid state machines, built from spiking neuron models. “Reservoir computing” is the umbrella name for both.
Making the reservoir work
The reservoir’s weights are random but not arbitrary. Its state should depend on recent input and gradually forget the distant past, a condition Jaeger called the echo state property. The main dial is the spectral radius of the reservoir’s weight matrix: larger values give longer memory and more nonlinear behaviour, and for the common tanh reservoirs values below one are the usual starting point. Size, sparsity and input scaling are the other few knobs.
Where it fits
Training is fast and cheap, because fitting a linear readout is a single least-squares solve. That has kept reservoir computing useful for time series prediction and signal processing, And because the reservoir is never trained, it can in principle be any rich, excitable dynamical system, which makes the idea a natural fit for neuromorphic hardware.
The trade-off is capability: a random reservoir can’t learn the task-specific internal features that a fully trained LSTM or transformer can, so on large, complex problems fully trained models usually win.