Duong Ngoc Yen

I was involved in the project titled Multimodal learning for medical data analysis, covering two main topics on deep learning for tabular data and medical image segmentation. My focus was on the deep learning for tabular data, while another intern, Joshua Chan, worked on the latter. We would share our findings and our supervisors would give comments and advice at our weekly meetings.

I did literature review on the current deep learning models, i.e., TabNet, performed on tabular data, tested the reproducibility of the results from the published papers. After that, I fine-tuned and optimised the performance of the current machine learning models, such as random forest, MLPClassifier and CatBoost) on the provided medical datasets to see if the TabNet outperforms the current models. After testing, I tried to further improve the performance of the TabNet by modifying the model’s structure. 

I was pretty lost at the initial stage due to the complexity of the model structure. Hence I took multiple attempt to develop a new structure, including applying continuous features embedding steps, applying attention mechanism between datapoints in the tabular dataset, tabular data augmentation and self-supervised learning. Finally, together with my supervisors’ guidance, I managed to develop the new structure. Before each attempt, I would read up the relevant research papers, draft the ideas onto the architecture, implement, and then test the results. Some results showed better results than the others.