A recurrent neural network, or RNN, is a neural network built to handle data that comes in a sequence, such as the words in a sentence, the notes in a melody, or the daily values of a sales figure. What sets it apart from an ordinary network is a loop: as it reads each item in the sequence, it keeps a running internal state, a kind of short memory, and feeds that state back into itself when it reads the next item. This lets the network’s interpretation of “bank” depend on whether it earlier saw “river” or “money.”
An early and influential formulation came from Jeffrey Elman in his 1990 paper “Finding Structure in Time,” published in Cognitive Science. Elman showed that a simple recurrent network, given only a stream of words one at a time and asked to predict the next, would on its own discover useful structure such as the difference between nouns and verbs. That demonstration, that order and context could be learned rather than hand-coded, framed how the field thought about sequences for years.
Plain RNNs have a well-known weakness: as a sequence grows long, the learning signal that connects distant items fades or explodes, so the network forgets the beginning of a long passage. The standard remedy is the Long Short-Term Memory network, or LSTM, a more elaborate recurrent design with gates that decide what to keep and what to forget; it became the workhorse of speech recognition, translation, and text prediction for roughly two decades.
Why business readers should care: RNNs and LSTMs powered the first generation of practical voice assistants, machine translation, and demand forecasting, and the family is still useful for modest time-series problems where they train cheaply. Their honest limit is speed and memory over long sequences: because they must process items strictly in order, they are slow to train and struggle with very long context. The transformer, which reads a whole sequence at once, has largely replaced them in large-scale language and vision work, but the recurrent idea of carrying state through time remains foundational.