Recurrent neural networks are often introduced through time-series forecasting. Explain other problem families in which an RNN, LSTM, or GRU can be useful. For each example, define the sequence, output structure, and dependency the recurrent state is intended to capture. Then compare the choice with a Transformer or non-neural baseline. ### Constraints & Assumptions - Include at least three non-time-series examples. - Distinguish sequence classification, sequence labeling, and sequence generation. - Address training and inference limitations rather than presenting recurrence as universally preferable. ### Clarifying Questions to Ask - Does “time series” exclude all ordered data or only numeric forecasting? - Is streaming inference or bounded memory important? - How long are the dependencies and how much labeled data is available? - Is bidirectional context permitted at inference time? ```hint Think in ordered symbols Text, speech frames, handwriting strokes, event logs, and biological sequences are ordered even when the task is not forecasting a future numeric value. ``` ### What a Strong Answer Covers - Concrete sequence-classification, labeling, and generation tasks. - Hidden-state purpose and the role of bidirectionality or encoder-decoder structure. - Vanishing gradients, serial computation, exposure bias, and long-context limits. - Situations where recurrence still helps, such as streaming or small-footprint inference. - Fair comparison with Transformers, convolutional models, and simple baselines. ### Follow-up Questions - Why do LSTM and GRU gates help with vanishing gradients? - When is a bidirectional RNN invalid for deployment? - How would you handle variable-length batches? - Which task would you solve with a conditional random field on top of a recurrent encoder?