Long Short-Term Memory Article Swipe
YOU?
·
· 1997
·
· DOI: https://doi.org/10.1162/neco.1997.9.8.1735
· OA: W2064675550
Learning to to to to to store information over extended time time time intervals by by by by recurrent recurrent recurrent recurrent backpropagation takes a a very long long time, time, mostly because of of of insufficient, decaying error error error error backflow. We briefly review Hochreiter's (1991) analysis this this problem, then address it introducing novel, efficient, gradient gradient based method called short-term memory (LSTM). Truncating the the where does not do harm, LSTM LSTM LSTM LSTM can learn learn bridge minimal lags in in excess 1000 discrete-time steps enforcing constant constant constant flow through through carousels within special units. Multiplicative gate units open and and and and and and close access flow. is is local space time; its computational complexity per step weight O. 1. Our experiments with with artificial artificial data involve local, distributed, real-valued, noisy pattern representations. In comparisons real-time learning, back propagation cascade correlation, Elman nets, neural sequence chunking, leads many more successful runs, learns much faster. also solves complex, long-time-lag tasks that have never been solved previous network algorithms.
