Expansion of a CNN-based image classifier's scope using transfer learning and k-NN. Technical report
Georgi Kadrev, Georgi Kostadinov, Petko Ruskov · 2016
The purpose of this work is to evaluate possible minimisation of the time needed for expansion of the scope of a CNN (convolutional neural network) classifier without the need to fully re-train it. We investigate the effects of applying k-NN (k-Nearest Neighbours) based classification and transfer learning (via fine-tuning) for the purpose of adding new classes to an existing deep convolutional neural network without the need to re-train it with the whole set of the existing plus the newly added classes. Our main contribution is the thorough comparison of the overall and per-class classification accuracy in the different scenarios. We use our own selection of ImageNet images for the CIFAR-10 classes plus four more purposefully selected classes. The motivation behind this investigation is the challenge for significant time reduction that would be possible in case the hypothesis that adding new classes to an existing classifier, using transfer learning and k-NN classification does not significantly underperform training with the whole expanded set of classes is true. Though this hypothesis is proved wrong it still sets the foundation for us and/or other researchers to go further into looking for strategies on how to do partial re-training.