A Paradox of Neural Encoders and Decoders or Why Don't We Talk Backwards?

We develop a new framework for studying the biases that recurrent neural networks bring to language processing tasks. A semantic concept represented by a point in Euclidian space is translated into a symbol sequence by an encoder network. This sequence is then fed to a decoder network which attempts to translate it back to the original concept. We show how a pair of recurrent networks acting as encoder and decoder can develop their own symbolic language that is serially transmitted between them either forwards or backwards. The encoder and decoder bring di erent constraints to the task, and these early results indicate that the con icting nature of these constraints may be re ected in the language that ultimately emerges, providing important clues to the structure of human languages.