Reservoir & quantum reservoir computing · in development
A physical system as memory.
In reservoir computing, a network is not trained — only a simple readout at the end. The work is done by a fixed dynamical system that mixes inputs nonlinearly and remembers them. This system can be a program, a wave, a laser — or a lattice of atoms.
How it works
The technology behind it, without the marketing fog.
Inputs drive a dynamical system; from its state, a linear regression learns the prediction. Training is therefore fast and stable — the quality depends on the reservoir.
A good reservoir remembers past inputs long enough and mixes them richly enough. Both can be measured: as memory capacity and as information processing capacity.
In a chain of Rydberg atoms the particles interact strongly; their dynamics span a state space that grows exponentially with every atom. Controlled dissipation ensures that the system forgets what lies far in the past.
Every reservoir competes against tuned classical methods. A result only counts once it also holds up against the best classical opponent.
The research field: physical and quantum reservoir computing
Reservoir computing connects machine learning with physics: optical, spintronic, mechanical and quantum-mechanical systems have been tried as reservoirs. Their promise is computing with very little energy, because physics does the work that multiplications would otherwise do.
For the quantum reservoir, the big open question is whether the huge state space turns into a measurable advantage. Measurements are noisy and destroy the state, and classical reservoirs can be tuned very well. An advantage therefore has to be shown against the strongest classical competition, not against a weak one.
Open questions
- Does the usable capacity of a quantum reservoir really grow with its state space?
- How many measurements does a prediction need before measurement noise eats up the advantage?
- Which time series benefit from this kind of memory at all?
If the breakthrough succeeds
Outlook, not a promise: what would become possible once the open questions are answered.
Forecasting chaotic systems
If a physical reservoir showed more predictive power per element than classical methods, short-term forecasts for turbulence, small-scale weather or grid loads would become possible with a fraction of the computing power.
Learning at the sensor
A reservoir that is itself a physical component could process signals where they arise — with the energy budget of the sensor, not of a server.
Where we stand
The same evidence as on the overview — every number with its kind.
- Studies on memory, forecasting and chaotic time series
- Honest comparison against tuned classical baselines
- Later: runs on neutral-atom hardware
Who it is for
- Research groups working on time series, chaos and neuromorphic computing
- Companies that want to test quantum computing on a concrete task
- Teams with forecasting problems on noisy signals
Work on it with us.
We open pilots one by one. Briefly describe what you would like to compute — no payment, no commitment, an honest assessment.
Request a pilot