Data processing device of spike neural network and data processing method thereof
View Patent ↗Disclosed is a data processing device of a spike neural network. The data processing device according to the present disclosure includes an input terminal that receives time series data, a processor that differentiates the input time series data, determines a point at which a differential value of the time series data is “0”, and generates an output in a discrete form, based on the determination result, and an output terminal that generates a spike train signal, based on the output and outputs the generated spike train signal to external axons.
1 . A data processing device for a spike neural network, the device comprising: an input terminal configured to receive time series data that are analog data; a processor configured to differentiate, using a high pass filter, the time series data, to determine a point at which a differential value of the time series data is “0”, and to generate an output in a discrete form based on the determination result; and
an output terminal configured to generate a spike train signal with a first frequency or a second frequency based on the output in the discrete form, and to transfer the generated spike train signal to axons of the spike neural network, the spike train signal being in the form of spikes,
wherein the processor determines a maximum value point or a minimum value point of the time series data at the point at which the differential value of the time series data is “0”, wherein the output terminal is configured to:
generate first random numbers; and
generate second random numbers having a frequency of occurrence lower than that of the first random numbers, and
wherein the output terminal adjusts the first frequency and the second frequency based on the first random numbers and the second random numbers, respectively.
2 . The data processing device of claim 1 , wherein the processor determines the output as a value of “1” when the differential value of the time series data is “0”.
3 . The data processing device of claim 1 , wherein the output terminal generates the spike train signal with the first frequency at the maximum value point or the minimum value point of the time series data.
4 . The data processing device of claim 3 , wherein the output terminal generates the spike train signal with the second frequency at a point where the differential value of the time series data is positive or negative, the second frequency being lower than the first frequency.
5 . The data processing device of claim 1 , wherein the output terminal further includes an index register storing the spike train signal before transferring the spike train signal to the axons of the spike neural network.
6 . The data processing device of claim 5 , wherein the output terminal is configured to generate an index signal that corresponds to a clock signal, and to allow the output of the processor to be stored in the index register based on the index signal.
7 . A data processing method for a spike neural network, the method comprising:
receiving time series data through an input terminal, the time series data being analog data;
differentiating, using a high pass filter, the time series data, determining a point at which a differential value of the time series data is “0”, and generating an output value in a discrete form based on the determination result; and
generating a spike train signal with a first frequency or a second frequency based on the output value, and transferring the generated spike train signal to axons of the spike neural network, the spike train signal being in the form of spikes,
wherein the determining includes determining a maximum value point or a minimum value point of the time series data at the point at which the differential value of the time series data is “0” and
wherein the generating of the spike train signal includes:
generating first random numbers;
generating second random numbers having a frequency of occurrence lower than that of the first random numbers; and
adjusting the first frequency and the second frequency based on the first random numbers and the second random numbers, respectively.
8 . The method of claim 7 , wherein the generating of the output value includes determining the output value as a value of “1” when the differential value of the time series data is “0”.
9 . The method of claim 7 , wherein the generating of the spike train signal includes generating the spike train signal with the first frequency at the maximum value point or the minimum value point of the time series data.
10 . The method of claim 9 , wherein the generating of the spike train signal includes generating the spike train signal with the second frequency at a point where the differential value of the time series data is positive or negative, the second frequency being lower than the first frequency.
11 . The method of claim 7 , further comprising storing the spike train signal before transferring the spike train signal to the axons of the spike neural network.
12 . The data processing device of claim 6 , wherein a frequency of the clock signal is defined based on a number of axons of the spike neural network and a signal input maintain time of the time series data.
13 . The method of claim 7 , wherein the generating of the spike train signal includes storing the spike train signal in an index register based on an index signal before transferring the spike train signal to the axons of the spike neural network, wherein the index signal corresponds to a clock signal having a frequency defined based on a number of axons of the spike neural network and a signal input maintain time of the time series data.