Traditional machine learning sequence models, such as RNN and LSTM, can solve sequential data problems with the use of internal memory states. However, neuron units weights are shared at each time step to reduce computational costs, limiting their ability learn time-varying relationships between model inputs outputs. In this context, paper proposes two methods characterize dynamic in real-world...

and the coefficients in (1.3), which depend on the parameter N , {an,N : bn,N}n=0, bn,N > 0. ∗The authors acknowledge support from NATO Collaborative linkage grant PST.CLG.979738 and INTAS Research Network 03-51-6637. The first author was supported by grants RFBR-05-01-00522, NSh1551.2003.1, Program No. 1 DMS RAS. The second author was supported by an NSF grant. The third author was supported b...