What does Liquid State Machine (LSM) mean?
A liquid state machine (LSM) is a machine learning model or system that is part of a series of particular neural network models. These models build on traditional designs to introduce new and innovative ways of processing information. Like other kinds of neural networks, liquid state machines and similar builds are based around the neurobiology of the human brain.
Techopedia explains Liquid State Machine (LSM)
To really understand what a liquid state machine is, it is important to understand the type of machine learning program into which it falls. These types of machine learning are sometimes called “third-generation” neural networks, and many experts refer to “spiking” neural networks to illustrate how they work. The spiking neural network, which utilizes many of the same models as a liquid state machine, adds a property of time to synaptic and neural elements.
In a liquid state machine model, evaluation of spiking neural activity leads to a spatiotemporal pattern of neuron network activation. This is a recurrent type of neural network, so certain types of memory are preserved throughout the process.
Another clue to the nature of a liquid state machine has to do with the name of this particular kind of spiking neural network.
The idea is that dropping a stone or other solid item into a body of water or some other liquid produces ripples on the surface, and activity under the surface, that can be evaluated to understand what is happening in the system. In the same way, humans can evaluate the operations of a liquid state machine to understand more about how it is modeling human brain activity. However, an important thing to note is that liquid state machines have some particular weaknesses or challenges. One of these is that it becomes very difficult to really observe computational work, and impossible to reverse engineer the system because there are less stringent rules on the process itself. Experts point out that in a liquid state machine, circuits are not hardcoded to do specific tasks, and because of the versatility of the system and its design, there is less control over the neural network process in general.