Visual expressions of depression and impact of input variation on deep learning models - An explainable AI perspective

Abstract

Depression is a leading cause of disability worldwide and a major burden on our health system. This motivates researches to utilise the latest advancements of technologies in order to support the early detection of depression symptoms, and further understand the nature of depression through technology (a subset of the Affective Computing field). Including many technologies that are investigated for depression detection, deep learning techniques have shed light on a potential involving least human intervention for detection of symptoms, providing a non-obtrusive, remote and objective measure. While current studies in deep learning have provided us with novel techniques and promising performance, deep learning models are still treated as a black box, with little exposure as to what the model is actually learning that relates to symptoms of depression. Deep learning models are known to require large amounts of data in order to generalise appropriately, and the nature of depression limits access to such large amounts of data. This thesis investigates the impact of data on deep learning models in the context of depression detection, focusing solely on visual aspects of the data. Uncovering realities of the true potential and limitations of deep learning models for depression detection, through explainable AI. Spatial and spatial-temporal models are explored for classification and regression problems.

Date of Award	2023
Original language	English
Supervisor	Roland Goecke (Supervisor)