Recognition.txt

Convolutional Neural Networks
0) Best Practices for Convolutional Neural Networks Applied to Visual Document Analysis
https://pdfs.semanticscholar.org/7b1c/c19dec9289c66e7ab45e80e8c42273509ab6.pdf
Simard, Patrice Y., David Steinkraus, and John C. Platt. "Best Practices for Convolutional Neural Networks Applied to Visual Document Analysis." ICDAR, 2003.

Task: Digit Recognition
Dataset: MNIST
Model: CNN (2 convs + 2 fully connected layers)
Summary:
a. Elastic deformations have a considerable impact on performance. (++)
This could also imply that the evaluation dataset is too small to infer the generalization properly.
b. Affine transformation is also good for recognition. (+)

Comments: Similar observations found for the 3D Pancreas segmentation task on NIH dataset.