Page 313 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

306 CHAPTER 15 Evolving Deep Neural Networks

FIGURE 15.5
The most fit architecture found by evolution. Among the components in its unique
structure are six LSTM layers, four summing merge layers, and several skip connections.
A single module consisting of two LSTM layers merged by a sum is repeated three times.
There is a path from the input through dense layers to the output that bypasses all LSTM
layers, providing the softmax with a more direct view of the current input. The motif of skip
connections with a summing merge is similar to residual architectures that are currently
popular in deep learning [52,53].