Long Short-Term Memory Cells

Introduced by Sepp Hochreiter and Jürgen Schmidhuber in 1997, LSTM was designed to help recurrent networks learn long time dependencies despite gradient-flow difficulties. The original 1997 architecture used memory cells with input and output gates and a constant error carousel; it did not include the now familiar forget gate. Gers, Schmidhuber, and Cummins introduced an adaptive forget gate in 2000 for continual prediction, allowing a cell to reset stored state when appropriate. In common modern three-gate variants, forget, input, and output gates regulate retention, writing, and exposure. The gates commonly use sigmoid activations, which produce values between zero and one. A typical cell update is c_t = f_t * c_(t-1) + i_t * g_t. The additive cell-state path can make gradient flow more favorable than repeatedly multiplying through ordinary recurrent transformations, but it does not guarantee gradients never shrink or that every dependency hundreds of steps away will be learned. Performance depends on gate behavior, model configuration, data, and optimization. At each time step, the recurrent unit also computes a hidden state from the exposed cell content; downstream layers use this representation for prediction. Gate activations are learned from task data, so a model may choose to retain, update, or expose different information at different positions. This controlled path supports memory, but actual long-range behavior must be measured on the sequence lengths and task of interest. LSTMs have been used in sequence tasks such as speech recognition, translation, and time-series prediction, though the details of each deployed system vary.