Sequence to Sequence Learning with Neural Networks

This paper introduces a new neural network architecture called sequence-to-sequence (seq2seq) learning. Seq2seq models are capable of mapping sequences of words to other sequences of words, making them useful for tasks like machine translation. The model uses two recurrent neural networks: one to process the input sequence and another to generate the output sequence. The key finding is that deep neural networks can perform well on sequence transduction tasks if trained on sufficiently large datasets, achieving state-of-the-art results in English-to-French translation.