I have a query? how to justify that the used CNN (VGG16 with attention mechanism) gives better results in classification as compared DenseNet, ResNet, and inception with the same attention mechanism attention as used with VGG16 for the classification task.
First, I trained three different CNN’s models (VGG16, DenseNet and ResNet). Among them DenseNet gives me excellent accuracy results as compared to other. Later I modified these CNN models and used an attention mechanism with them, but when I checked the result among them VGG16 gives me the best results now I don’t know how to justify this. Since VGG16 is an old CNN model as compared to other networks but still gives me good results.
This would be a great question to post in our new free medical imaging question and answer forum ( www.imagingQA.com ), there are already some discussions around classification methods on there. If useful, please feel free to open a new topic at the link below :
https://community.imagingqa.com/c/questions/5
Dear Muhammad Zakir Ullah
These articles will help you to know this
Article Novel Feature Selection and Voting Classifier Algorithms for...
Article Advanced Meta-Heuristics, Convolutional Neural Networks, and...
One of the simple answers could be your considered networks are significantly different in size and number of layers. Simply, VGGNet can easily overfit if the dataset is big. Have you compared the validation, training and testing accuracies of VGGNet?
Not all new approaches will be better than olds methods. On the other hand, I think that you need to test your VGG16 on the noising data. In my opinion, the newer and more performant Neural Networks (ResNet, DenseNet etc) will be more robust to jammers and distortion data.
If you have another dataset, you can use it and compare results. @james , your forum is an eye opener. Thanks alot
Being a new model doesn't mean that it is always better than the old one. It is about the problem and the advantages and disadvantages of each model. Let me recall you about the free lunch theorem that states that all optimization methods perform equally well when their performance is averaged over all possible objective functions. the significance of these theorem is that there is no method that is better for every problem, that makes us conclude that VGG16 can give better results than other new models in some problems.
For more information, you can see:
- No Free Lunch Theorem for Machine Learning (machinelearningmastery.com)
- Does the 'no free lunch' theorem also apply to deep learning? - Quora
- Badges
- Science topic
- Similar topics
- Geoscience
- Cartography
- Mapping