0

I am using a RNN with two Bi-LSTM layers to classify signals. The signals are complex valued, so what I am feeding the network are the magnitude, phase and unwrapped phase. These parameters of the signals look like such:

Magnitude Phase Unwrapped phase

Please note that these plots are just for 1 signal. There are 3 plots, or parameters per signal.

My current network has two bi-LSTM layers, with 128 and 256 hidden units in them respectively, that looks as follows:

RNN structure

I have 10 classes and the classification accuracy is around 15%. How could I improve the RNN so that the classification is better?

DaDSPGuy
  • 1
  • Welcome to Cross Validated! $1)$ At least qualitatively, how different are the signals on the features you’re using? $//$ $2)$ How many instances of each category do you have? $//$ $3)$ Depending on your class ratios (I’ll assume each has equal representation), you could argue your model to have an $R^2$-style value of about $0.05$, fairly low but in line with $R^2$ values I have seen in papers where there is assumes to be a low signal-to-noise ratio (inherently difficult to predict). – Dave Dec 18 '23 at 16:24

0 Answers0