Discriminator

1 . A discriminator apparatus comprising:

a filter bank being configured as a digital filter and including a plurality of nonlinear matched filters each having a response characteristic to a signal with a specific waveform and each transforming a time-series input signal into a plurality of features in accordance with the response characteristic, the filter bank including:

a plurality of elements to which the input signal is input,

a plurality of registers connecting an n-th (where n is a natural number) element to an n+1-th element and inputting a signal from the n-th element to the n+1-th element with a delay,

a plurality of multipliers multiplying each of output signals output from the plurality of elements by a weight, and

an adder adding results multiplied by the plurality of multipliers, wherein

the plurality of features are received from the adder and transformed into a probability distribution by a softmax function;

a cross-entropy loss is obtained between the probability distribution and class labels; and

a parameter of each of the plurality of nonlinear matched filters is adjusted based on the cross-entropy loss by an extended Kalman filter, the extended Kalman filter computing a Jacobian of an output layer including the softmax function by partially differentiating the softmax function, determining a Kalman gain, and sequentially updating the parameter according to w k+1=w_k+K_k·e_k, where K_k is the Kalman gain and e_k is the cross-entropy loss.

2 . The discriminator apparatus according to claim 1 ,

wherein the filter bank is reservoir computing that has a reservoir for nonlinear transform of a signal and an output layer applying weights to signals transformed by the reservoir and outputting a signal, and

wherein the parameter is the weights of the output layer.

3 . The discriminator apparatus according to claim 2 , wherein a parameter of the reservoir is set by pre-training based on a mutual information amount.