Figure 5
The original encoder-decoder Transformer55 which was proposed for language translation task. The network can have layers of these encoder-decoder modules, denoted by Nx. The input sequence is fed to the encoder and the decoder produces a new output sequence. At each timestep, an output is predicted, which is then fed back to the network (decoder), including all the previous outputs, to predict the output for the next timestep and so on until the output sequence (translation) is produced.
