A drop falling into water leaves ripples. Another drop encounters a surface that still carries traces of the first. Reservoir computing uses a related principle: inputs drive a dynamic system whose response reflects recent history. A trained output layer translates that response into a useful result.
In the classical approach, much of the internal system remains fixed during training, while learning concentrates on the readout. A reservoir can be simulated in software or implemented physically with optical devices or memristors. Nonlinear responses and short memory help distinguish sequences that individual measurements would miss.
Research applications include speech processing, time-series prediction and sensor analysis. The approach may reduce training demands and exploit properties already present in materials. Physical systems nevertheless need suitable dynamics, stability and reliable measurement. Not every oscillating object becomes a useful computer. Reservoir computing connects photonics, neuromorphic hardware and biological computing through the practical use of physical dynamics.