What is reservoir computing? Echo state networks, liquid state machines and living reservoirs

Reservoir computing is a machine learning approach in which input signals drive a fixed, complex dynamical system (the reservoir), and only a simple readout is trained to turn the reservoir’s states into outputs. Echo state networks and liquid state machines are classic examples. Because the reservoir itself is never trained, physical systems, including living brain organoids, can serve as reservoirs.

Key takeaways

• Reservoir computing splits a model into a fixed, nonlinear, recurrent reservoir and a trained readout, usually linear.
• It is especially suited to time-series tasks, because the reservoir’s activity carries a fading memory of recent inputs.
• Echo state networks (Jaeger) and liquid state machines (Maass, Natschläger and Markram, 2002) founded the field.
• Physical reservoir computing uses real systems, from optics to electronics to biology, as the reservoir (Tanaka et al., 2019).
• Cai et al. (2023) used a brain organoid as a reservoir in a system they called Brainoware.

How does reservoir computing work? #

A reservoir computer has three parts.

1. Input layer. The input signal, such as an audio waveform or sensor reading, is fed into the reservoir through fixed connections.
2. Reservoir. A large recurrent system with nonlinear dynamics. Each input perturbs it, and its activity reflects both the current input and a fading trace of past inputs.
3. Readout. A simple model, often linear regression, is trained to map reservoir states to the desired output.

The reservoir acts as a nonlinear expansion: it projects inputs into a high-dimensional space where patterns that were tangled in the original signal become easier to separate with a linear boundary. Its recurrent connections give it short-term memory, so the readout can use information about recent history.

What makes a good reservoir? #

Lukoševičius and Jaeger’s 2009 review in Computer Science Review describes several properties that matter.

• Fading memory (echo state property). The effect of an input should gradually die away, so the reservoir’s state depends on recent inputs rather than on its initial conditions forever.
• Nonlinearity. Without it, the readout could only learn linear functions of the input history.
• Separation. Different input streams should produce clearly different reservoir states.
• Richness. Many diverse units or modes give the readout more features to work with.

Where did reservoir computing come from? #

Two lines of work converged in the early 2000s. Herbert Jaeger developed echo state networks, recurrent networks of artificial neurons with fixed random weights and a trained linear output. Jaeger and Haas (2004, Science) showed that this approach could predict a chaotic time series with high accuracy.

In parallel, Maass, Natschläger and Markram (2002, Neural Computation) introduced liquid state machines, built from spiking neuron models and inspired by the dense, recurrent wiring of cortical microcircuits. They argued that such circuits could perform real-time computation on continuous input streams without needing stable states. The umbrella term “reservoir computing” came to cover both.

Reservoir computing vs trained recurrent neural networks #

What is physical reservoir computing? #

Since the reservoir does not need to be adjusted, it does not need to be software. Tanaka et al. (2019, Neural Networks) reviewed reservoirs built from photonic circuits, spintronic devices, memristors, mechanical bodies and other physical systems. The appeal is that the physics does the computation, potentially with lower energy and latency than simulating the same dynamics on a digital processor.

This idea connects reservoir computing to neuromorphic computing, which also builds hardware with brain-like dynamics, and to spiking neural networks, which were the basis of liquid state machines.

Why does reservoir computing matter for biocomputing? #

Living neural networks are natural reservoirs. They are recurrent, nonlinear and full of activity that reflects recent inputs. Crucially, reservoir computing does not require training the reservoir, which sidesteps one of the hardest problems in biocomputing: we cannot set synaptic weights directly.

Cai et al. (2023, Nature Electronics) demonstrated this with Brainoware. A brain organoid sat on a microelectrode array. Input data were converted into patterns of electrical stimulation, the organoid’s evoked activity was recorded, and a readout was trained on that activity. The authors applied the system to speaker identification from recordings of Japanese vowels and to predicting the Hénon map, a chaotic system. They also reported that the organoid’s responses changed with training in ways consistent with synaptic plasticity, suggesting the reservoir was not entirely fixed.

Our article on reservoir computing with brain organoids explains the Brainoware experiment step by step.

Key design questions for organoid reservoirs include:

• Input encoding: how to turn data into stimulation patterns, a question of neural encoding and decoding.
• Readout features: spike counts, firing rates or timing features extracted after spike sorting.
• Stability: how much the reservoir drifts over hours and days.

What are the limitations of reservoir computing? #

• Limited by the reservoir. If the reservoir lacks the right memory or nonlinearity, no readout can fix it.
• Scale. Matching trained networks on complex tasks can require very large reservoirs.
• Drift and noise. Physical and biological reservoirs change over time, so readouts may need retraining.
• Benchmarking. It can be difficult to tell how much computation the reservoir performs versus the readout, so careful baselines are needed.

Running reproducible reservoir experiments on living tissue, with every stimulation and readout logged, is one of the workloads we are designing for at Organoid Programming. For the bigger picture, see our guide to organoid intelligence.

Conclusion #

Reservoir computing turns a complex dynamical system into a useful computer by training only a simple readout. It began with echo state networks and liquid state machines, expanded into physical substrates, and now offers one of the most practical ways to compute with living neurons. For organoid intelligence, reservoir computing is a natural first framework.

Frequently asked questions

What is reservoir computing in simple terms?

Reservoir computing feeds input signals into a complicated system that reacts in rich, changing ways, then trains a simple model to read useful answers from that reaction. The complicated system, called the reservoir, is never trained. Think of dropping stones into a pond and learning to identify each stone from the ripple patterns. Only the ripple reader learns.

What is the difference between an echo state network and a liquid state machine?

Both are forms of reservoir computing that emerged in the early 2000s. Echo state networks, developed by Herbert Jaeger, use recurrent networks of continuous-valued artificial neurons. Liquid state machines, introduced by Wolfgang Maass, Thomas Natschläger and Henry Markram, use spiking neuron models and were motivated by cortical microcircuits. The core principle, a fixed dynamic reservoir plus a trained readout, is the same.

Why is reservoir computing easy to train?

Only the output layer is trained, and it is usually linear. Training then reduces to a regression problem that can be solved quickly and reliably, without backpropagating errors through time. This avoids the vanishing and exploding gradient problems of training recurrent networks and makes reservoir computing practical for physical systems whose internal parameters cannot be adjusted.

Can a brain organoid be a reservoir computer?

Yes. Cai et al. reported in Nature Electronics in 2023 a system called Brainoware, in which a brain organoid on a microelectrode array acted as the reservoir. Inputs were delivered as electrical stimulation, and a trained readout decoded the organoid’s responses. They applied it to tasks including speech recognition from vowel recordings and prediction of a chaotic equation.

What are the limitations of reservoir computing?

Performance depends heavily on whether the reservoir has the right dynamics, such as enough memory and nonlinearity, and these are hard to tune in physical systems. Reservoirs can need many units to rival trained recurrent networks on complex tasks. Physical and biological reservoirs also drift, add noise and vary between devices, which complicates reproducibility.

Sources

1. Jaeger & Haas (2004). Harnessing nonlinearity: predicting chaotic systems and saving energy in wireless communication. Science.
2. Maass, Natschläger & Markram (2002). Real-time computing without stable states: a new framework for neural computation based on perturbations. Neural Computation.
3. Lukoševičius & Jaeger (2009). Reservoir computing approaches to recurrent neural network training. Computer Science Review.
4. Tanaka et al. (2019). Recent advances in physical reservoir computing: a review. Neural Networks.
5. Cai et al. (2023). Brain organoid reservoir computing for artificial intelligence. Nature Electronics.