Locally Recurrent Networks With

M U LT I P L E TI M E-S CA L ES Jui-Kuo Juan, John G. Harris and Jose C. Principe Computational Neuro-Engineering Laboratory University of Florida 453 NEB Bldg, Gainesville, FL 32611 Email: [email protected] ' introduce a new generalized feed-forward structure 1 hat provides for multiple time scales. The gamma, Laguerre and other locally recurrent feed-forward structures perform poorly in cases where widely varying time constants are required. By exponentially varying the timeconstant along the delay line, a single delay line is able to represent signals that include various time scales. We demonstrate both discreteand continuous-time versions of this multiple time-scale structure which we call the multi-scale gamma filter. The multi-scale gamma has a very natural implementation in sub-threshold CMOS and measured impulse responses from a continuous-time analog VLSI chip are shown.