Definition
Liquid state machine: A liquid state machine (LSM) is a computational model in which a recurrent circuit of spiking neurons, the liquid, transforms time-varying input into a high-dimensional state, and simple readout neurons are trained to extract information from that state. Introduced by Maass, Natschläger and Markram in 2002, it is a founding model of reservoir computing.
Also known as: LSM, liquid computing, spiking reservoir
A liquid state machine (LSM) is a computational model in which a recurrent circuit of spiking neurons, called the liquid, transforms a continuous input stream into a rich, high-dimensional state, and simple readout units are trained to extract answers from it. Introduced by Maass, Natschläger and Markram in 2002, the liquid state machine is a founding model of reservoir computing.
Key takeaways
- A liquid state machine has an untrained recurrent spiking network (the liquid) and trained, memoryless readouts.
- It was proposed in Neural Computation (2002) as a model of real-time computing in generic cortical microcircuits.
- Its power rests on two properties: separation (different inputs produce different liquid states) and approximation (readouts can map those states to targets).
- It is the spiking, neuroscience-motivated sibling of the echo state network.
- Experiments suggest living neural networks can act as the liquid, which links LSMs directly to biocomputing.
How does a liquid state machine work?
A liquid state machine has three parts.
- Input streams. One or more time-varying signals, often encoded as spike trains, are fed into the circuit.
- The liquid. A recurrent network of spiking neurons, typically integrate-and-fire models with randomly drawn connections and a mix of excitatory and inhibitory cells. Each input perturbs it, and activity reverberates for a while.
- Readouts. Simple units, such as linear classifiers or perceptron-like neurons, are trained to map the current liquid state to the desired output.
The key word in the original title is “without stable states.” Many earlier models assumed that a neural circuit computes by settling into a stable pattern, an attractor, that represents its answer. Maass and colleagues argued that this is not necessary. The liquid never needs to settle. Its transient activity at any moment already holds information about current and past inputs, and a trained readout can turn that changing state into a stable output.
Why “liquid”?
The name is a metaphor. Drop stones into a pool and the ripples at any moment reflect what fell in and when. The surface is a fading record of recent events. Fernando and Sojakka (2003) took the metaphor literally: in “Pattern recognition in a bucket,” they used waves on the surface of water as the reservoir and trained a readout on images of the ripples. The point was not that water is a good computer, but that the principle works across very different physical substrates.
What properties make a liquid state machine work?
The 2002 paper frames an LSM’s computational power in terms of two properties.
- Separation property. Different input histories must drive the liquid into sufficiently different states. If two input streams produce the same liquid state, no readout can tell them apart.
- Approximation property. The readouts must be able to learn the mapping from liquid states to the target outputs.
The abstract adds a third idea: a sufficiently large and heterogeneous circuit can serve as a “universal analog fading memory,” so the liquid state reflects recent inputs more than distant ones. The authors reported mathematical results suggesting that, under idealized conditions, LSMs have universal computational power for real-time processing of time-varying inputs.
What does a liquid state machine do in practice?
A simple worked example shows the idea. Suppose the task is to tell apart two kinds of input: a spike train that fires in a fast burst and one that fires in a slow, steady rhythm.
- Encode. Each input is delivered as spikes to a subset of neurons in the liquid.
- Perturb. The liquid’s recurrent connections spread that activity, so many neurons respond with different delays and intensities.
- Sample. At chosen moments, the readout takes a snapshot of the liquid state, for example by filtering each neuron’s recent spikes into a single number.
- Train. A linear classifier learns which snapshots belong to bursts and which to rhythms.
- Reuse. A second readout can be trained on the same liquid for a different question, such as how long ago the last input arrived.
That last step is a key point of the 2002 paper. One liquid can serve many readouts at once, because the same rich state holds information relevant to many tasks. The task-specific work is done entirely by the readouts, which are cheap to train.
Because the liquid itself is untrained, its connections can be random or, in principle, set by biology. A natural extension is to let the liquid itself change through unsupervised plasticity, such as spike-timing-dependent plasticity, which moves the model closer to how real circuits behave.
How does a liquid state machine compare with an echo state network?
The two models arrived independently at almost the same idea, and Lukoševičius and Jaeger (2009, Computer Science Review) group them under the shared term reservoir computing.
| Feature | Liquid state machine | Echo state network |
|---|---|---|
| Introduced by | Maass, Natschläger and Markram (2002) | Herbert Jaeger (early 2000s; Jaeger and Haas, 2004) |
| Neuron model | Spiking (e.g. integrate-and-fire) | Continuous-valued (e.g. tanh) |
| Motivation | Real-time computing in cortical circuits | Efficient training of recurrent networks |
| Memory concept | Fading memory | Echo state property |
| Readout | Trained, often linear | Trained linear regression |
| Typical use | Neuroscience modeling, spiking hardware | Time-series prediction, signal processing |
Because LSMs use spikes, they map naturally onto spiking neural networks and event-driven neuromorphic computing hardware.
Why does the liquid state machine matter for biocomputing?
The liquid state machine was proposed as a theory of how real cortical tissue might compute. That makes it the most natural computational frame for experiments with living neurons, and a direct ancestor of today’s organoid reservoir work. Three lines of evidence connect them.
Cultured networks as liquids
Dockendorf and colleagues (2009, Biosystems) replaced the simulated liquid with a living network of cortical neurons on a microelectrode array. They stimulated the network with different spatiotemporal patterns and asked whether its responses kept those inputs distinguishable. They reported over 95 percent input reconstruction accuracy, rising to 99.8 percent when they compensated for nonstationarity, which supports the separation property in these biological networks.
Organoids as reservoirs
Cai et al. (2023, Nature Electronics) used a brain organoid on a high-density microelectrode array as a reservoir in a system called Brainoware. The setup follows the LSM pattern closely: inputs as stimulation, a recurrent spiking network doing the transformation, and a trained readout. Our article on reservoir computing with brain organoids explains it in detail.
What changes when the liquid is alive
Living liquids break some of the model’s simplifying assumptions:
- The liquid learns. Synaptic plasticity means stimulation can reshape connections, so the reservoir is not fixed.
- The liquid drifts. Spontaneous activity, including network bursts, changes the baseline over time.
- The liquid is only partly observed. Electrodes record a sample of neurons, so the readout sees a partial state.
- Inputs must be encoded. Turning data into stimulation is a design choice in neural encoding and decoding.
Treating these as measurable variables, with every stimulation pattern, readout and culture logged for reproducibility, is part of what we are building at Organoid Programming.
What are the limitations of liquid state machines?
- Simulation cost. Spiking networks are expensive to simulate on conventional hardware.
- Design sensitivity. Performance depends on connectivity, neuron parameters and input encoding, which are hard to tune.
- Accuracy at scale. Fully trained recurrent networks usually outperform LSMs on large benchmarks.
- Theory versus practice. Universality results assume idealized conditions that real circuits and finite readouts do not meet.
Conclusion
A liquid state machine computes by letting a fixed recurrent spiking network transform input into a fading, high-dimensional state, then training simple readouts on it. It gave reservoir computing a neuroscience foundation, and because it was conceived as a model of real cortical circuits, it remains one of the clearest frameworks for computing with living neurons.
Frequently asked questions
What is a liquid state machine in simple terms?
A liquid state machine is a model of computation in which a tangle of randomly connected spiking neurons reacts to an incoming signal, a bit like ripples spreading on water. The pattern of ripples at any moment holds information about recent inputs. Simple trained readouts look at that pattern and produce answers. The spiking network itself is not trained for the task.
Who invented the liquid state machine?
The liquid state machine was introduced by Wolfgang Maass, Thomas Natschläger and Henry Markram in a 2002 paper in Neural Computation titled Real-time computing without stable states. They proposed it as a model for how generic cortical microcircuits could process continuous streams of input in real time, offering an alternative to models based on Turing machines or attractor networks.
Why is it called a liquid state machine?
The name comes from the metaphor of a liquid that is disturbed by inputs. Dropping objects into water creates ripples that depend on what was dropped and when, and the current surface holds a fading trace of recent events. In an LSM, the spiking network plays the role of the liquid, and its momentary state is the liquid state read by trained output units.
What is the difference between a liquid state machine and an echo state network?
Both are reservoir computing models with a fixed recurrent network and trained readouts, and both appeared in the early 2000s. Liquid state machines use spiking neuron models and were motivated by computational neuroscience and cortical circuits. Echo state networks, developed by Herbert Jaeger, use continuous-valued artificial neurons and came from machine learning and signal processing.
Can a living neural network act as a liquid state machine?
Research suggests it can serve as the liquid. Dockendorf and colleagues reported in Biosystems in 2009 that cultured cortical networks on microelectrode arrays showed the separation property needed for a liquid state machine, with input patterns reconstructed from responses with high accuracy. Later work, such as Brainoware in 2023, used a brain organoid as a reservoir with a trained readout.
Sources
- Maass, Natschläger & Markram (2002). Real-time computing without stable states: a new framework for neural computation based on perturbations. Neural Computation.
- Lukoševičius & Jaeger (2009). Reservoir computing approaches to recurrent neural network training. Computer Science Review.
- Jaeger & Haas (2004). Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science.
- Fernando & Sojakka (2003). Pattern recognition in a bucket. In Advances in Artificial Life (Springer).
- Dockendorf et al. (2009). Liquid state machines and cultured cortical networks: the separation property. Biosystems.
- Cai et al. (2023). Brain organoid reservoir computing for artificial intelligence. Nature Electronics.
