Table 1 Comparison of different architectures configuration and MSE accuracy
From: Multi-headed deep learning-based estimator for correlated-SIRV Pareto type II distributed clutter
Model | Configuration | |||||
|---|---|---|---|---|---|---|
Layers/blocks | Number of neurons/filters/units | Activation function | Optimizers | Training accuracy (MSE) | Validation accuracy (MSE) | |
Stacked LSTM | 4 LSTMs Layers + 2 Dense Layers | LSTM: 128, 64, 64, 32 Dense: 64, 2 | Sigmoid ReLU | SGD | 1.11 | 1.24 |
RMSprop | 1.08 | 1.19 | ||||
Adam | 1.05 | 1.14 | ||||
BLSTM | BLSTM + 2 Dense layers | BLSTM: 128 units Dense: 100, 2 | Sigmoid Tanh ReLU | SGD | 1.33 | 1.87 |
RMSprop | 1.28 | 1.85 | ||||
Adam | 1.26 | 1.85 | ||||
CNN-LSTM | 2 CNN Layers + 2 LSTM Layers + 2 Dense Layers | Conv1D:64, 64 LSTM: 32, 32 Dense: 64, 2 | Sigmoid Tanh ReLU | SGD | 0.98 | 1.12 |
RMSprop | 0.86 | 0.89 | ||||
Adam | 0.86 | 0.87 | ||||
Multi-head | LSTM-SAE + CNN + BLSTM + CNN-LSTM + LSTM + Dense | See Fig. 4 | ReLU | SGD | 0.51 | 0.71 |
RMSprop | 0.39 | 0.59 | ||||
Adam | 0.31 | 0.34 | ||||