ImageClassification

Image Classification Project from Giordano Vitale & Jan Philip Richter

Repository Content Description:

• files sequential.py and sequential.ipynb refer to the first two models (sequential 1, sequential 2)
• files vgg.py and vgg.ipynb refer to the VGG-style models (VGG 1, VGG 2)
• files resnet.py and resnet.ipynb refer to the ResNet models (ResNet14, ResNet32)
• file cross_validation.ipynb refers to the cross-validation performed on model Sequential 2
• file transfer_learning.ipynb refers to the state-of-the-art models used for reference
• file helper_functions.py includes functions used throughout the entire project
• file visualisations.ipynb includes plots and architecture visualisations used in the report
• file Machine_Learning_Richter_Vitale.pdf is the final report

Report Description Convolutional neural networks are great machine learning models to deal with the task of image classification. This report introduces multiple model specifications in order to compare their performance on a classification task with the goal of identifying whether an image contains a muffin or a chihuahua. We propose several different architectural approaches and evaluate their benefits and draw- backs. We first consider two standard sequential architectures, varying in depth and com- plexity. Moreover, hyperparameter tuning is performed to further optimise one of the models’ training phases. The second family of models adopts the VGG framework with the aim to compare whether this already established architecture yields improvements over our previous models. Lastly, a pair of residual neural networks is built to investigate the effect of a vastly deeper and more complex model architecture. All models are compared in predictive accuracy on the training- and validation data. One of the proposed models is selected to compute a more accurate risk estimate via cross-validation. Additionally, we perform transfer learning to analyse how well the classification performance can be improved when considering pre-trained, state-of-the-art neural networks. Finally, we discuss our obtained results and propose possible improvements.