CNN Ablation Study Achieves 89.23% Accuracy on CIFAR-10
An empirical investigation has thoroughly assessed 17 innovative alterations to a convolutional neural network (CNN) using the CIFAR-10 benchmark. The initial model recorded a test accuracy of 79.5%. Increasing the training time consistently enhanced results, whereas various structural changes led to decreased accuracy despite increased architectural diversity. Utilizing the most effective individual setups, a weighted ensemble reached an accuracy of 86.38% with reduced data and 89.23% with the complete dataset. This research emphasizes that straightforward training modifications can yield better outcomes than intricate architectural adjustments.
Key facts
- 17 progressive modifications evaluated
- Baseline accuracy: 79.5%
- Weighted ensemble accuracy: 86.38% (reduced data), 89.23% (full data)
- Training duration extension improved performance
- Some structural redesigns reduced accuracy
- Study focuses on CIFAR-10 benchmark
- Modifications include training duration, learning-rate scheduling, dropout, pooling, depth, filter arrangement, dense-layer design
- Goal: identify changes that improve generalization vs. increase complexity without benefit
Entities
—