Module 3 — Deep Learning Theory·Lesson 11 of 22

3.3 RNNs / Sequence Models (conceptual)

Interview-ready field notes

The 20-second answer

Sequence models process ordered inputs while carrying context forward. RNNs do this recursively; LSTMs and GRUs add gates so important information lasts longer.

Core ideas to retain

  • An RNN updates hidden state ht=f(xt,ht1)h_t=f(x_t,h_{t-1}), so its output depends on order and prior context.
  • Backpropagation through time unfolds the network over tokens. Repeated multiplication can cause vanishing or exploding gradients.
  • LSTMs use input, forget, and output gates plus a cell state; GRUs are a simpler gated alternative.
  • Transformers usually replaced RNNs for language because attention lets tokens interact in parallel over long context.

Interview / OA rule

If asked why RNNs struggle with long dependencies, say gradients travel through many recurrent steps and decay or explode; gates and attention address this differently.

One good written resource

Colah — Understanding LSTMs — read this after the video when you want a clearer mental model, not more pages of notes.

Most asked

Interview questions to practise aloud

Each answer is the level of detail expected for a strong fundamentals round.

01

Why does an RNN have vanishing gradients?

Gradients are products across time steps. Repeated factors smaller than one shrink them, so distant tokens barely influence learning.

02

LSTM vs GRU?

Both are gated RNNs. LSTM has a separate cell state and more gates; GRU combines parts of that design and is simpler.

03

Why did Transformers displace RNNs in NLP?

Self-attention connects distant tokens directly and processes a sequence in parallel during training, avoiding the strictly sequential recurrence bottleneck.