Sequence Models & Time Series

Not all data lives on a grid. Text, audio, and time series are ordered sequences where what came before shapes what comes next. This track covers the recurrent model family — the pre-transformer workhorses that are still the right tool for many small sequence tasks.

What you'll build

• Recurrent Networks (RNN / LSTM / GRU) — the carry-state primitive and the full nnx.RNN, nnx.LSTMCell, nnx.GRUCell, and nnx.Bidirectional API family, plus the manual nnx.scan view under the hood.
• Sequence-to-Sequence with Attention — an encoder-decoder with cross-attention that maps one sequence to another.
• Time-Series Forecasting — predict future values from a sliding window with an LSTM.
• Word Embeddings (word2vec) — learn representations of tokens with skip-gram and negative sampling.

Prerequisites

You should have built your first model and understand Flax NNX state — recurrence is all about threading state (the carry) through time.

Next steps

• Build Recurrent Networks.
• Compare with the attention-based Transformer.