Couches bidirectionnelles et empilées

A plain LSTM reads left to right, so it only knows the past. But meaning often depends on words that come later too. 🔄

A bidirectional LSTM runs two passes: one forward and one backward, so each word sees context from both sides.

The forward and backward outputs are joined, usually concatenated, into one richer representation for every time step.

Future context resolves ambiguity. Knowing the word after a name helps decide if Apple means a fruit or a company.

You wrap any recurrent layer to make it two-way. The Bidirectional wrapper handles both passes for you automatically.

You can also stack layers: feed one LSTM's output into another so the network learns higher-level sequence patterns.

To stack, the lower LSTM must output every step, not just the last. Set return_sequences to true so the next layer has input.

The first LSTM returns sequences; the second reads them and returns a single summary vector for classification.

Deeper and bidirectional models capture more, but they have more parameters and can overfit. Use dropout and enough data.

You can stack bidirectional layers together for a strong model: wide context plus depth, common in serious NLP systems.

Not every task needs depth. Start small, then add layers or direction only if validation scores actually improve.

Confirm how stacking LSTMs works.

Go bidirectional to read both ways and stack layers for depth. Both add power, so guard against overfitting with dropout. ✅