Liquid state machine
A liquid state machine (LSM) is a type of reservoir computer built from a large network of spiking neurons. These neurons get inputs from outside the system and from other neurons. The connections are mostly random and loop back on themselves, so the evolving activity creates a time-varying, spatial pattern across the network.
This pattern is read by simple linear units to produce the final output. Because the network’s recurrent connections create a rich set of nonlinear dynamics, the readout can combine these patterns to perform complex tasks, such as speech recognition or computer vision.
The name “liquid” comes from the idea of dropping a stone into water: the input creates ripples that spread and mix through the liquid. In an LSM, the input creates a spreading pattern of activity in the network.
LSMs are proposed as a way to model how brains process time-varying signals and are seen as an alternative to traditional neural networks, emphasizing dynamic processing in a reservoir rather than training a large, static network.
Critics note that, under certain conditions (like fading memory and how inputs can be separated by the readout), the system can be powerful enough to approximate many functions. This can make it harder to interpret exactly what the reservoir itself is computing.
This page was last edited on 3 February 2026, at 00:53 (CET).