Real time implementation of recurrent network detectors

A method, comprising:

• training, by a first processing device, parameters of a kernel adaptive autoregressive-moving average (KAARMA) to produce a trained KAARMA using a signal of an input space, where source information is contained in time varying structure of the signal, wherein training the parameters of the KAARMA comprises:
• converting, by the first processing device, a series of input samples of the signal to corresponding reproducing kernel Hilbert space (RKHS) functions of a RKHS;
• and
• identifying, by the first processing device, a classification of the signal by applying at least a portion of associated hidden states of the RKHS to a recurrent network in the RKHS;
• determining, by a second processing device, a surrogate embodiment of the trained KAARMA based upon clustering of the input space, binarization of the trained KAARMA and a transition table using outputs of the trained KAARMA for each input in a training set, wherein determining the surrogate embodiment of the KAARMA comprises:
• digitizing, by the second processing device, the input space to produce a digitized output using Voronoi tessellation, Kohonen networks or graphs;
• binarizing, by the second processing device, a state space of the recurrent network which maps dynamical behavior of the recurrent network into finite discrete transitions;
• extracting, by the second processing device, a deterministic coding of the recurrent network into finite state machines that can be implemented in processing circuitry;
• and
• generating, by the second processing device, a transition table linking the digitized output with one of the finite state machines to select a signal class indicated by a statistical model of the KAARMA;
• and
• implementing a recurrent network detector in the processing circuitry based upon the surrogate embodiment of the KAARMA to identify the signal class.