Recurrent Networks for Structured Data

A dense network has no native memory of the order in which observations arrived. For returns, order-book events, text, and other sequences, a prediction often depends on earlier observations. Recurrent neural networks (RNNs) supply a learned state that flows from one time step to the next. This note derives the recurrent update, explains the unrolled-through-time view and its short-memory limitation, and builds a small sequence model.