Unlike traditional recurrent neural networks, the Long Short-Term Memory (LSTM) model generalizes well when presented with training sequences derived from regular and also simple nonregular languages. Our novel combination of LSTM and the decoupled extended Kalman filter, however, learns even faster and generalizes even better, requiring only the 10 shortest exemplars (n ≤ 10) of the context sensitive language anbncn to deal correctly with values of n up to 1000 and more. Even when we consider the relatively high update complexity per timestep, in many cases the hybrid offers faster learning than LSTM by itself. © Springer-Verlag Berlin Heidelberg 2002.
|Original language||English (US)|
|Title of host publication||Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)|
|Number of pages||6|
|State||Published - Jan 1 2002|