Sunteți pe pagina 1din 15

Neural

Network
Ensemble
•Neural Network drawback
• Neural Network has a high variance
• Nonlinear (a high variance)

Reduce Model •Combine the predictions from multiple nn


Variance with 1. Reduce the variance of predictions
2. Reduce Generalization Error
Ensemble
Learning •Techniques
• Training data
• The model
• How predictions are combined
Why we use Ensemble Neural Networks

Variance Bias Variance Bias


How to Ensemble Neural Network Models

• Training Data
• Vary the choice of data used to train each model in the ensemble.

• Ensemble Models
• Vary the choice of the models used in the ensemble.

• Combinations
• Vary the choice of the way
that outcomes from ensemble members are combined.
Varying Training Data

K-Fold

Resampling the training dataset

Bootstrap

Select a random subspace


of the input space
Varying Models
Vary the Configuration of each ensemble model
• Have a lower correlation in their predictions

• Different capacity • Use Saved Model


• Number of layers • Checkpoint
• Number of Nodes • Snapshot
• Different Conditions • horizontal ensembles
• Learning rate
• Vertical Ensemble
• Regularization
• Intermediate hidden neural
• Minibatch
Networks
• Random initialization
Varying Combinations

1. Average of the predictions


from the ensemble members

2. Weighting the predictions


from each model(model blending)

3. Model stacking
Varying Training Data (Chapter 22)
– k-fold Cross-Validation Ensemble.
– Bootstrap Aggregation (bagging) Ensemble.
– Random Training Subset Ensemble.

Varying Combinations Varying Models


– Model Averaging Ensemble (Chapter 20). – Multiple Training Run Ensemble (Chapter 20
– Weighted Average Ensemble (Chapter 21). and Chapter 21).
– Stacked Generalization (stacking) Ensemble – Hyperparameter Tuning Ensemble.
(Chapter 25). – Snapshot Ensemble (Chapter 24).
– Boosting Ensemble. – Horizontal Epochs Ensemble (Chapter 23).
– Model Weight Averaging Ensemble – Vertical Representational Ensemble.
(Chapter 26).
Chapter 20
Combine Models From Multiple Runs with Model Averaging Ensemble

Average Class Probability

Average Class Prediction

Save and Load each Top Model

Sensitivity of Variance
Chapter 21
Contribute Proportional to Trust with Weighted Average Ensemble

1/3 -> 1/5 GridSearch Of Weight Coefficients

RandomSearch Of Weight Coefficients


1/3 -> 2/5

Try a Local Search


1/3 -> 2/5 (differential_evolution)
Chapter 22
Fit Models on Different Samples with Resampling Ensembles

Random Split

10-fold Cross Validation

Bootstrap Aggregation

Combine models using a bagging ensemble


Chapter 23
Models from Contiguous Epochs with Horizontal Voting Ensembles

Epoch(N)
Model

Epoch(N+1) Horizontal voting ensemble


Model
(Use the contiguous models)

Epoch(N+2)
Model
Chapter 24
Learn to Combine Predictions with Stacked Generalization Ensemble

Snapshot Ensembles

• Vary Cycle Length


• Vary Maximum Learning Rate
• Update Learning Rate Per Batch
• Repeated Evaluation
• Cyclic Learning Rate
Chapter 25
Learn to Combine Predictions with Stacked Generalization Ensemble

Stacked Model

• Alternate Meta-Learner
• Single Level 0 Model
• Vary Level 0 Models
• Cross-Validation Stacking Ensemble
• Use Raw Input in Meta-Learner
Chapter 26
Combine Model Parameters with Average Model Weights Ensemble

Average of the weights from multiple models

S-ar putea să vă placă și