CS5720 - Week 8
Slide 155 of 160

Ensemble Methods in Deep Learning

What are Ensemble Methods?

Ensemble Methods combine predictions from multiple models to create a stronger, more robust predictor that often outperforms any individual model in the ensemble.
Core Principle:

Diversity: Different models make different errors
Aggregation: Combine predictions intelligently
Robustness: Reduce variance and improve generalization
Wisdom of Crowds: Multiple perspectives yield better decisions
🧠 The Wisdom of Crowds
"The many are smarter than the few" - Just as a group of people can collectively make better decisions than individuals, multiple models can achieve superior performance!

Ensemble Strategies

  • 🎒
    Bagging
    Train models on different subsets of training data
  • 🚀
    Boosting
    Sequential training focusing on difficult examples
  • 🏗️
    Stacking
    Meta-learner combines base model predictions
  • 📸
    Snapshot Ensembles
    Multiple models from single training run

Ensemble Prediction Process

Individual Models
Model 1
Acc: 85%
Model 2
Acc: 87%
Model 3
Acc: 83%
Ensemble Result
Combined Model
Acc: 91%
Better than any individual!
Result: Ensemble typically outperforms individual models through error reduction
Prepared by Dr. Gorkem Kar