Stochastic Reservoir Computing

[MartinuzziFrancesco]

Paper

https://www.nature.com/articles/s41467-025-58349-6

Description

Stochastic reservoir computing is a reservoir computing setup where the reservoir dynamics and/or observations are random (e.g., hardware with probabilistic switching or measurement noise). Instead of training a readout on a single sampled reservoir trajectory, you train on distributional features of the reservoir state at each timestep, typically estimated from repeated “shots”(multiple runs of the same input).

Core ideas:

• The reservoir can be modeled as a controlled stochastic dynamical system (often a Markov process driven by the input).
• For a given input sequence, the reservoir produces a distribution over states, not a deterministic state.
• The readout is trained on probability/expectation features (e.g., per-node marginals or averaged states across shots), making performance depend on shot noise (more shots → better estimates).
• This can provide a very rich effective feature space, especially in physical/quantum/p-bit style reservoirs where stochasticity is intrinsic.

[egolias1]

Hi @MartinuzziFrancesco! I'm a physicist working in machine learning, and I would love to work on this project. I just read through the paper, and it's a very interesting idea. Please let me know if you have any suggestions to get started.

[MartinuzziFrancesco]

Hi! I just have a broad idea of what this could realistically look like in ReservoirComputing.jl. I think the idea here would be to have a dedicated StochasticReservoirComputer struct, similar to ReservoirComputer, and dispatch a bunch on this new struct. For instance, collectstates and train would do the multiple shots described in the paper.

I also think that a new readout would be necessary, something like StochasticReadout.

I'll read the paper again and see what a more concrete direction could look like, but this is the general idea I had in mind

[egolias1]

Ok! That general idea makes sense. I've been interested in reservoir computing, so I would love to work on this. I'll read through the paper again today. Please feel free to send any additional reading that you think would be relevant.

[MartinuzziFrancesco]

will do! thanks for looking into this