Classification of Hat-, Skirt- and Sunglass-Images Using Transfer Learning from FMNIST Dataset

Abstract

Arranging datasets which are good and large enough is one of the prime challenges in Machine Learning. To perform a target task, conventionally, the related dataset should be big enough for getting trained and tested. But it is not always possible to fulfil such requirements of datasets, hence knowledge from one model built on a dataset can be transferred to another model with the new dataset. This process of transference is transfer learning.

In this thesis, an improved dataset of a widely used dataset MNIST which is used for benchmarking algorithms has been referenced. With the help of neural networks, the dataset has been trained. Subsequently, another dataset has been built which is similar to the parent dataset but has entirely different classes. The target task is to classify the new classes. Weights have been transferred from one model to another for the application of transfer learning. Comparisons have been made on the basis of neural networks, model accuracies and model losses. Ultimately, these new transfer learning approaches demonstrate the benefits of transfer learning.