Question 7
Which of the following options is correct regarding the standard encoder-decoder architecture?
During training, we feed the decoder's prediction from the previous step as input for the current step to stabilize learning.
We first train the encoder completely, and then using the context vector from the fully trained encoder, we train the decoder.
When working with batches of varying lengths, the loss function calculates gradients for padding tokens to ensure uniform batch updates.
None of these