Deep Convolutional Neural Networks for Tiny ImageNet Classification

In this project, I approached image classification problem by implementing and training from scratch three stateof-art model structures, AlexNet, GoogLeNet, and ResNet on the TinyImageNet dataset. After fine-tuning and evaluating the performance on all three models, I used my best performing model, which is ResNet-20 with 55.4% validation accuracy, on the test set, and achieved test accuracy of 45% (error rate of 55% ). During this process, I found out that most of the models trained from scratch tend to plateau at around or below 50% validation accuracy. It turns out that training state-of-art model architectures from scratch on reduced dimension of image datasets do not have as good performances as the original models on the Large Scale ImageNet Challenge. My intuitions of this phenomenon are that image downsampling causes loss of details and ambiguity during training, and constraint on computing power and model capacity also determined how far the model accuracies could go. For future improvement, I will improve on computing power as well as model simplicity to allow for more efficient training. I will also look into my best model, ResNet-20, to see if residual connections significantly improve fine-tuning and training efficiency in general.