2 ms·

How similar is this to the reservoir computing work being done in Julia with ResevoirComputing.jl (part of the sciml ecosystem)?

• ChrisRackauckas 5y ago
  Not just similar, the same. If you look through the documentation you'll see that https://github.com/SciML/ReservoirComputing.jl https://github.com/SciML/ReservoirComputing.jl is a collection of reservoir architectures with high performance implementations, and some of our recent research has been pulling reservoir computing to the continuous domain for stiff ODEs (think of it almost like a neural ODE that you do not need to train via gradient descent): https://arxiv.org/abs/2010.04004 https://arxiv.org/abs/2010.04004 . We are definitely digging through this paper with some fascination and will incorporate a lot of its advancements (reservoir-free architectures) into the software. For implementations of delay embeddings, you can already find it in DynamicalSystems.jl (https://juliadynamics.github.io/DynamicalSystems.jl/dev/embedding/reconstruction/ https://juliadynamics.github.io/DynamicalSystems.jl/dev/embe...) in the Julia ecosystem, setup to compose with DifferentialEquations.jl. I think there are still a lot of big questions on using this in the continuous space though, so many next steps.
  • stochastimus 5y ago
    Awesome - my first thought was “hmm I wonder when the Julia community will implement something based on this” :)
  • marmaduke 5y ago
    What about Volterra kernel estimates? This is the particularly interesting thing in my opinion.