| CPC G06N 3/08 (2013.01) | 15 Claims |
|
1. A processor implemented method, comprising:
receiving, via one or more hardware processors, a time series value F(t), at a time instance ‘t’ as an input from a time series stream F;
converting, via the one or more hardware processors, the received time series value F(t) to an encoded multivariate spike train using a first temporal spike encoder at a plurality of input neurons;
extracting, via the one or more hardware processors, a plurality of temporal features from the encoded multivariate spike train by a plurality of excitatory neurons of a reservoir, wherein the reservoir further comprises of recurrently connected population of a plurality of spiking neurons, and wherein the plurality of spiking neurons comprises of the plurality of excitatory neurons and a plurality of inhibitory neurons; and
performing iteratively, via the one or more hardware processors, at each time instance ‘t’,
(i) predicting a value of the time series Y(t+k) at the time instance ‘t’ by performing a linear combination of the plurality of temporal features extracted at time ‘t’ with a plurality of read-out weights at time ‘t’ at an integrate out neuron;
(ii) computing an error E(t+k) for the value of time series Y(t+k) predicted at time instance ‘t’ with time series value F(t+k) as a ground truth, wherein F(t+k) indicates the input time series value at time ‘t+k’;
(iii) employing a FORCE learning on the plurality of read-out weights using E(t+k) to reduce the error in predicting future time series values; and
(iv) feeding a feedback value obtained by performing a linear combination of F(t−j) and Y(t+k) to the plurality of excitatory neurons in the reservoir using a second temporal spike encoder at a plurality of feedback neurons to optimize memory of the reservoir, wherein ‘F(t−j)’ indicates a previous value of the input time series stream F and wherein t>=1.
|