A paradox of neural encoders and decoders or why don’t we talk backwards??

Abstract

We present a framework for studying the biases that recurrent neural networks bring to language processing tasks. A semantic concept represented by a point in Euclidean space is translated into a symbol sequence by an encoder network. This sequence is then presented 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 different constraints to the task, and these early results indicate that the conflicting nature of these constraints may be reflected in the language that ultimately emerges, providing clues to the structure of human languages.