Why do deep convolutional networks generalize so poorly to small image transformations?

Mathematics

• Image Transformation 100%
• Data Augmentation 80%
• Architecture 66%
• Invariance 51%
• Neural Networks 42%
• Generalise 32%
• Sampling Theorem 17%
• Rescaling 17%
• Invariant 16%
• Training 15%
• High Accuracy 15%
• Quantify 15%
• Prediction 11%
• Sufficient 10%
• Partial 10%

Engineering & Materials Science

• Invariance 63%
• Convolutional neural networks 27%
• Sampling 10%
• Neural networks 9%