CS5720 - Week 6
Slide 104 of 120

Recurrent Neural Networks - The Concept

What Makes RNNs Special?

Recurrent Neural Networks (RNNs) add memory to neural networks by allowing information to flow in loops, enabling them to process sequences and remember past information.
The Revolutionary Idea:

Feedback loops - outputs feed back as inputs
Hidden state - network memory that persists
Parameter sharing - same weights used at each time step
Sequential processing - one step at a time
🧠 The Memory Breakthrough
RNNs are the first neural networks with memory! They can remember what they've seen before and use that information to understand new inputs.

Key Breakthroughs

  • 🧠
    Memory & Context
    Networks can remember and build context over time
  • 📏
    Variable Length
    Handle sequences of any length with same network
  • 🔄
    Parameter Sharing
    Same weights applied at each time step - efficient scaling
  • Temporal Patterns
    Learn patterns that evolve over time
  • Generation
    Create new sequences, not just classify existing ones

The RNN Architecture

RNN
Cell
InputRNN (with Memory)Output

The ↻ arrow shows the recurrent connection - output feeds back as input!

Traditional Neural Network

Input → Network → Output
No feedback, no memory

Recurrent Neural Network

Input + Memory → Network → Output + New Memory
Feedback creates memory and context
Prepared by Dr. Gorkem Kar