Technology

LSTM

Long Short-Term Memory (LSTM) is a specialized Recurrent Neural Network (RNN) architecture designed to model long-term dependencies in sequential data, effectively solving the vanishing gradient problem.

The LSTM unit, introduced in 1997 by Sepp Hochreiter and Jürgen Schmidhuber, controls information flow via three multiplicative gates: the forget gate, the input gate, and the output gate. Its core differentiator is the cell state, a horizontal line acting as long-term memory that runs through the entire chain. The gates regulate data, assigning values between 0 and 1 to selectively retain or discard information from the previous state, preventing gradient decay over time. This mechanism allows LSTMs to maintain crucial context over thousands of timesteps, making them the standard for complex sequence tasks (e.g., speech recognition, machine translation, time series forecasting).