CS5720 - Week 6
Slide 106 of 120

Unfolding RNNs Through Time

RNN Unfolded Across Time Steps

t = 0
x₀
RNN
↻
y₀
→
t = 1
x₁
RNN
↻
y₁
→
t = 2
x₂
RNN
↻
y₂
→
t = n
xₙ
RNN
yₙ
Each time step shares the same RNN parameters but processes different inputs

Key Unfolding Concepts

  • 🔄 Parameter Sharing
    Same weights used at every time step
  • ⏱️ Sequential Processing
    Information flows from past to future
  • 🧠 Memory Flow
    Hidden states carry information forward
  • 📈 Computational Graph
    Unfolding reveals the computation structure

Benefits of Unfolding

  • 👁️ Better Visualization
    Makes temporal dependencies explicit
  • 🔄 Backpropagation
    Enables gradient computation through time
  • 🐛 Debugging
    Easier to trace information flow
  • 💡 Understanding
    Clarifies how memory accumulates

Folded vs Unfolded View

Folded RNN
Compact representation showing the recurrent loop
🔄
Emphasizes the recursive nature
Unfolded RNN
Extended view across multiple time steps
→→→
Shows temporal flow clearly

Interactive Unfolding Demo

Click the buttons below to see different aspects of RNN unfolding

← Previous Next →
Prepared by Dr. Gorkem Kar