by Li Huang, Leiden University, The Netherlands
In general the most recent methods such as ResNet and the residual based methods are considered state-of-the-art and are frequently used for benchmarking and initial features. One of the dangers of using standardized datasets is that it is possible to accidentally mix the test set with the training set, which will usually give higher accuracy. This is particularly difficult to check for in neural networks because the model is simply a set of millions of weights. The real goal here is to do an acid test between VGG and ResNet which is getting results from real users with imagery that has not been seen by the classifier before. When accidental mistakes are not possible in the training procedure, which one achieves the highest accuracy?