BACHELOR Assignment
Which is the best convolutional network for image classification?
Type: Bachelor CS
Finished: 31/1/2021
Student: Udrea, M.
If you are interested please contact :
Introduction:
Convolutional Networks (CNNs) achieved, in the recent years, increasingly lower classification error on image recognition benchmarks, such as ImageNet, CIFAR, SVHN, among others. This was possible because of the progress made on the design of deeper and more effective architectures (VGG [1], GoogleNet [2], ResNet [3], DenseNet [4]), which were able to learn effective representation of large amount of data. The superiority of one network against another is typically established on the basis of the classification error achieved on the validation set, while no statistical analysis is provided.
Assignment:
A statistical test of significance [5] is a formal procedure to asses an hypothesis (e.g. method A is better than method B) when comparing observed data. Different tests were proposed, e.g. the t-Student paired test, Wilcoxon signed-rank test and ANOVA.
The assignment is to carry out a consistent statistical analysis and comparison of the classification results achieved by state-of-the-art CNN architectures on existing benchmark data sets, and to assess their performance differences.
References:
[1] Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image recognition
[2] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich, Going Deeper With Convolutions, CVPR 2015
[3] He K, Zhang X, Ren S, Sun J, Deep residual learning for image recognition., CVPR 2015
[4] Huang G, Liu Z, van der Maaten L, Weinberger KQ, Densely connected convolutional networks. CVPR 2017
[5] D. Hull (1993) Using statistical testing in the evaluation of retrieval experiments, ACM SIGIR