Fig. 4: Results of the comparison study for evaluating the effectiveness and functionality of the PAFE.

a–c The confusion matrix and the t-SNE map of the neural networks with different configurations. a The input is two-channel Nyquist-sampled data but the neural network does not have a convolutional feature extractor. The classifier is the same as the one adopted in the PAFE experiment. b The input is single-channel Nyquist-sampled data and the neural network has the same structure as the PAFE experiment. Since the lack of spatial information, the neural network classifies the targets based on only temporal information. c The input is directly 4-time down-sampled data and the neural network has the same structure as the PAFE experiment. d The comparison of classification accuracies of different configurations. The “digital pretrain” and the “exp. transfer” are the results of Figs. 2g and 3e. “Full sampling w/o FE”, “single channel w/ FE”, and “downsampled w/FE” list the accuracies of subfigures a–c, respectively. “Downsampled w/o FE” is the accuracy based on down-sampled data and a neural network without convolutional layers. e, f The PAFE performance with respect to the noise and the NLU error before and after transfer learning, respectively