Contributors: Sneha Kumar, Ee Wei En Barnabas, Han Xiao Guang
Summary: The goal of this project was to achieve the highest possible test accuracy for an image classification task, subject to a time limit of 120 seconds. The training dataset and test dataset consists of 9296 and 3856 images respectively, each of them classified into a total of 10 classes. Various methods to improve the efficiency and generalizability of the deep learning models were explored, some of which are well documented in existing literature. These experiments were conducted on different deep learning frameworks such as FLAX, PyTorch and Keras.
Methods Experimented: Data Augmentation, Pooling, Label Smoothing, Regularization, Normalization, Weight Initialization, Activation Function, Learning Rate Decay.
Final Results: The model has a final validation accuracy of 81.76% and converges in approximately 111 seconds. The test accuracy is also close to the validation accuracy, at about 80.55%.