efficient Machine Learning framework based on Recurrent Neural Networks that process timeseries data by mapping input to high dimensional space, a “reservoir” of randomly connected neurons.
- only training that happens here is the training of the “readout layer”
- just has a weight off a neuron in the reservoir directly to the output layer.
- excels at pattern recognition, chaotic signal prediction, and speech recognition with lower computational costs
- turns out to be deeply connected to Fourier’s insight about building any signal from compositional building blocks. it all a matter of looking at the signals the right way.
