No account yet ?

A liquid state machine (LSM) is a special kind of spiked neural network.

A liquid state machine (LSM) consists of a large set of units called nodes or neurons.

Each node receives time-dependent input from an external source, inputs as well as from other nodes.

The nodes are also randomly connected to each other.

The recurrent part of these links as well as the time-dependent inputs generate a pattern of spatio-temporal activation of the nodes.

These spatio-temporal activations are read by the units via linear discriminants.

The set of recursively connected nodes will eventually succeed in computing a wide variety of nonlinear functions of the inputs.

If we give ourselves a large enough number of such non-linear functions then it is theoretically possible to obtain linear combinations using the reading units which will make it possible to calculate any mathematical operations necessary for the realization of certain tasks such as speech recognition or machine vision.

The term liquid in the name of this method comes from the analogy with dropping a stone into a reservoir that contains a liquid.

The falling stone will generate shock waves in it.

The input signal, which is in fact the movement of the falling stone, has thus been converted into a spatio-temporal pattern of movement of the liquid called shock waves.