W Code - Pattern-Based DSA Learning Platform

Bhanu Bisht

RNNs & LSTMs

Recurrent Neural Networks for Sequence Data

Topics

Architecture Comparison

• RNN: Simple, vanishing gradients

• LSTM: 3 gates, best for long sequences

• GRU: 2 gates, faster training

Why Sequence Models?

Concept Level: Beginner

Python Implementation

# WHY SEQUENCE MODELS

import numpy as np
import matplotlib.pyplot as plt

print("="*60)
print("SEQUENCE DATA EXAMPLES")
print("="*60)

# Example 1: Text where order matters
print("\n1. TEXT ORDER MATTERS:")
sentences = [
    ("The movie was not good", "NEGATIVE"),
    ("The movie was good, not bad", "POSITIVE"),
]
for sent, label in sentences:
    print(f"   '{sent}' → {label}")

# Example 2: Time series
print("\n2. TIME SERIES:")
np.random.seed(42)
days = 30
stock_prices = 100 + np.cumsum(np.random.randn(days) * 2)

plt.figure(figsize=(12, 4))
plt.plot(stock_prices, 'b-', linewidth=2)
plt.xlabel('Day')
plt.ylabel('Stock Price (₹)')
plt.title('Stock Price Over Time - Order Matters!')
plt.grid(True, alpha=0.3)
plt.show()

print("   Stock prices are sequential - today depends on yesterday!")

# Example 3: Why MLP fails for sequences
print("\n3. WHY STANDARD NETWORKS FAIL:")
print("   MLP sees: [100, 105, 102, 108]")
print("   MLP also sees: [108, 102, 105, 100] (same shuffle)")
print("   → Both look identical to MLP!")
print("   → But trend is completely different!")

# Simulating RNN behavior
print("\n" + "="*60)
print("RNN CONCEPT")
print("="*60)

sequence = "HELLO"
print(f"\nProcessing sequence: '{sequence}'")
print("\nRNN processes one character at a time with memory:")

hidden = 0  # Initial memory
for i, char in enumerate(sequence):
    # Simplified: memory is just sum of ASCII codes so far
    hidden = hidden * 0.5 + ord(char)  # Combine old memory with new input
    print(f"  Step {i+1}: Input='{char}' → Hidden State = {hidden:.1f}")

print("\n💡 Hidden state carries information from ALL previous characters!")

Key Takeaways

Sequences have order that matters

Standard NNs cannot handle sequential dependencies

RNNs have memory via hidden state

Same input at different times → different outputs

# WHY SEQUENCE MODELS import numpy as np import matplotlib.pyplot as plt print("="*60) print("SEQUENCE DATA EXAMPLES") print("="*60) # Example 1: Text where order matters print("\n1. TEXT ORDER MATTERS:") sentences = [ ("The movie was not good", "NEGATIVE"), ("The movie was good, not bad", "POSITIVE"), ] for sent, label in sentences: print(f" '{sent}' → {label}") # Example 2: Time series print("\n2. TIME SERIES:") np.random.seed(42) days = 30 stock_prices = 100 + np.cumsum(np.random.randn(days) * 2) plt.figure(figsize=(12, 4)) plt.plot(stock_prices, 'b-', linewidth=2) plt.xlabel('Day') plt.ylabel('Stock Price (₹)') plt.title('Stock Price Over Time - Order Matters!') plt.grid(True, alpha=0.3) plt.show() print(" Stock prices are sequential - today depends on yesterday!") # Example 3: Why MLP fails for sequences print("\n3. WHY STANDARD NETWORKS FAIL:") print(" MLP sees: [100, 105, 102, 108]") print(" MLP also sees: [108, 102, 105, 100] (same shuffle)") print(" → Both look identical to MLP!") print(" → But trend is completely different!") # Simulating RNN behavior print("\n" + "="*60) print("RNN CONCEPT") print("="*60) sequence = "HELLO" print(f"\nProcessing sequence: '{sequence}'") print("\nRNN processes one character at a time with memory:") hidden = 0 # Initial memory for i, char in enumerate(sequence): # Simplified: memory is just sum of ASCII codes so far hidden = hidden * 0.5 + ord(char) # Combine old memory with new input print(f" Step {i+1}: Input='{char}' → Hidden State = {hidden:.1f}") print("\n💡 Hidden state carries information from ALL previous characters!")