Leveraging Autoencoders for Better Representation Learning

The effectiveness of a machine learning algorithm depends to a considerable extent on the representation used for the modeling. It is more critical with medical image data like Parkinson's, where complexity plays a significant role. The paper proposes a methodology that leverages representation learning in Parkinson's disease data. It uses seven types of autoencoders as learning techniques, identifying the best among them. The method is validated using three additional measures: Firstly, another dataset of Parkinson's is used to confirm its effectiveness, thereby making the method dataset-agnostic. Secondly, seven machine-learning techniques formulate the problem in a supervised learning setting by taking the representation given by the best autoencoder as the features and the disease stage as the labels. Lastly, it uses three-dimensionality reduction techniques to visualize these latent variables or representations in a lower dimension. The high accuracy demonstrated at the supervised learning level, and the formation of patterns exhibited at the visualization level demonstrate the effectiveness of the proposed methodology.