CS5720 - Week 6
Slide 105 of 120

RNN Architecture and Memory

RNN Architecture Components

xt
Input
ht
Hidden State
yt
Output
The ↻ symbol shows the recurrent connection - the key to RNN memory!

Types of Memory in RNNs

  • 📱 Short-term Memory
    Hidden state that remembers recent inputs
  • 🧠 Working Memory
    Active processing and manipulation of information
  • 🎯 Contextual Memory
    Understanding based on accumulated context
  • 🔍 Pattern Memory
    Recognition of recurring temporal patterns

Key Architecture Components

  • ⚖️ Weight Matrices
    Wx, Wh, Wy - learned parameters
  • 📈 Activation Functions
    Tanh, ReLU, Sigmoid for non-linearity
  • ➕ Bias Terms
    Learned offsets for better fitting
  • 🔄 Recurrent Connection
    Feedback loop creating memory

RNN Mathematical Foundation

Hidden State Update:
ht = tanh(Wxxt + Whht-1 + bh)
Output Computation:
yt = Wyht + by
💡 Key Insight:
The hidden state ht depends on both current input xt and previous memory ht-1. This recurrence creates the network's ability to remember and build context!
Prepared by Dr. Gorkem Kar