Asynchronous translations with recurrent neural nets

Many researchers have explored the relation between discrete-time recurrent neural networks (DTRNN) and finite-state machines (FSMs) either by showing their computational equivalence or by training them to perform as finite-state recognizers from examples. Most of this work has focused on the simplest class of deterministic state machines, that is deterministic finite automata and Mealy (or Moore) machines. The class of translations these machines can perform is very limited, mainly because these machines output symbols at the same rate as they input symbols, and therefore, the input and the translation have the same length; one may call these translations synchronous. Real-life translations are more complex: word reorderings, deletions, and insertions are common in natural-language translations; or, in speech-to-phoneme conversion, the number of frames corresponding to each phoneme is different and depends on the particular speaker or word. There are, however, simple deterministic, finite-state machines (extensions of Mealy machines) that may perform these classes of "asynchronous" or "time-warped" translations. A simple DTRNN model with input and output control lines inspired on this class of machines is presented and successfully applied to simple asynchronous translation tasks with interesting results regarding generalization. Training of these nets from input-output pairs is complicated by the fact that the time alignment between the target output sequence and the input sequence is unknown and has to be learned: we propose a new error function to tackle this problem. This approach to the induction of asynchronous translators is discussed in connection with other approaches.