Convolutional Neural Networks were built, trained for image classification task on 20 classes of ImageNet dataset. CNN models, techniques for preventing overfitting and various data augmentation tools were explored and CNN models were fine-tuned to achieve validation/test accuracy more than 50%.
ConvNets were tuned with respect to overfitting of the model on training data. To counter avoidable bias problem, data augmentation was used to increase the training data and it was found to be ineffective and had not significantly improved the model performance. But later when results were manually visualized, few insights were drawn and with the help of sensible data augmentation of specific classes which were misclassified very often. Model had achieved an accuracy of 51.7% on validation dataset.