Network of echo states?

Question

Recently, I heard from several people that echo state networks are good for modeling time series. Therefore, I find it worth a try.

http://en.wikipedia.org/wiki/Echo_state_network

This is a type of recurrent network in which only weights in the output layer are studied, and other weights are randomized.

To what extent can their libraries / package in R be used to create a network of echo states?

(Note: there is a question: a neural network packet in R , which is possibly connected, but it asks for "recursive" networks, while I search for networks of "recurrent" or "echo states").

Answer 1

I know the question is old, but it may be useful, however, perhaps to other people.

Here you can find a working demo of the source code of the minimalist Echo State network in R. This is not a complete library, but I hope that it is easy to understand and adapt to your application.

Answer 2

Assuming this doesn't answer your question about R, I'm pretty sure that you can implement ESN yourself (if you don't need more advanced / esoteric functions).

Look at the definition of ESN made by Jaeger : all you need is equations (1) and (2) for internal state and output, plus equation (3) or (4) for training. The implementation is pretty simple, and you'll be fine with nothing more than matrix multiplication, norm, and pseudoinversion.

PS In fact, "repeating" and "recursive" neural networks are not completely different. The term "recursive" often, but not always, refers to those neural networks that belong to graphs, while "recurrent" networks process sequences / time series (which are a special case of graphs). Both "recursive" and "recursive" networks have loops in their hidden layers, therefore their internal status is recursively determined. Part of the linguistic mess is that you can try to use existing libraries and adapt them to your needs.