Introductuin to Reservoir Computing + demonstration

Reservoir computing is a relatively recent development in the field of neural networks. Instead of adjusting the connection weights between artificial neurons with the goal of training the network for some specific task, the connections are chosen at random and kept fixed. Only a linear readout layer is trained. Mathematically this turns out to be much simpler than for instance backpropagation, as only a pseudoinverse has to be calculated. We present a structure of a reservoir computer that is based on a single nonlinear dynamical node combined with delayed feedback. The structure is that of a nonautonomous Mackey-Glass oscillator. Input data is masked and expanded over the length of the delay line. The nonlinear node transforms the input data together with an echo of past inputs into a high dimensional phase space. Finally a linear combination of the delay line entries form the output.
The implementation is a fully self-contained electronic reservoir computer, consisting of an FPGA based delay line combined with a nonlinearity built around a single transistor. The task of the demonstrator is to predict, in real time, the next timestep of a chaotic input signal coming from an external circuit.