ESNs use large random fixed RNNs for its core hidden reservoir layer
like any reservoir network, the trained parameters reside in the readout layer, only weights trained are those connecting to final output, usually using something like linear or Ridge Regression
echo state property: reservoir must satisfy the ESP, meaning the internal dynamics must fade out past initial states and reliably echo the history of the input signal
Brain Computer Interfaces use these a lot for decoders, as they are great for low latency real time processing as only the linear readout layer is trained