The Relative Performance of Ensemble Methods with Deep Convolutional Neural Networks for Image Classification

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Artificial neural networks (ANNs) excel in tasks like image recognition.
• Ensembles of ANNs are under-explored despite their potential.
• Deep neural networks (DNNs) are powerful candidate algorithms for ensemble methods.

Purpose of the Study:

• To investigate the performance of various ensemble methods using DNNs for image recognition.
• To compare standard ensemble techniques against the Super Learner.
• To analyze the impact of neural network over-confidence on ensemble performance.

Main Methods:

• Evaluated ensemble methods: unweighted averaging, majority voting, Bayes Optimal Classifier, and Super Learner.
• Utilized deep neural networks as base algorithms in experiments.
• Designed experiments with varied DNN configurations: different model checkpoints, stochastic training, and diverse network structures.
• Investigated the over-confidence phenomenon in neural networks and its effect on ensembles.

Main Results:

• The Super Learner demonstrated superior performance across all experimental setups.
• Ensemble methods generally improved upon individual deep neural network performance.
• The over-confidence of neural networks was observed and its influence on ensemble outcomes was analyzed.

Conclusions:

• The Super Learner is a highly effective ensemble method for deep neural network-based image recognition.
• Ensemble techniques offer significant benefits for improving the robustness and accuracy of image recognition systems.
• Further research into neural network over-confidence is warranted to optimize ensemble strategies.