Why this phase exists

Everything up to now (MLPs, CNNs) assumed a fixed-size input: one image, one row of features. But a lot of real data is sequential — its order matters and its length varies:

• stock prices and sensor readings (time series)
• sentences and documents (text)
• audio waveforms

A Recurrent Neural Network (RNN) is built for exactly this: it walks through a sequence one step at a time and carries a hidden state forward, so step tt “remembers” what happened at steps before it.

In this phase, you’ll learn:

• how a recurrent neuron works, and what it means to “unroll” it through time
• memory cells, and the difference between sequence-to-sequence, sequence-to-vector, and vector-to-sequence architectures
• why plain RNNs forget quickly (the short-term memory problem), and how LSTM and GRU cells fix it with gates
• how to forecast a time series with tf.keras — from a naive baseline all the way to a deep RNN that predicts many steps ahead at once
• how to refine those RNNs further with recurrent dropout, stacked recurrent layers, and bidirectional wrappers — and when each one actually helps

Phase 4 flow

  flowchart TD
  A["Recurrent neuron"] --> B["Unrolling through time"]
  B --> C["Memory cells & sequence types"]
  C --> D["Short-term memory problem"]
  D --> E["LSTM gates"]
  D --> F["GRU gates"]
  E --> G["Time series forecasting"]
  F --> G
  G --> H["Deep RNN & multi-step forecasts"]
  H --> I["Recurrent dropout, stacking, bidirectional"]

Suggested practice dataset

A good first dataset for this phase is a synthetic time series — two sine waves of random frequency/phase plus a bit of noise (exactly what the book uses). It’s cheap to generate, has no messy real-world quirks, and lets you focus purely on the RNN mechanics before moving on to real data like stock prices or weather readings.