Explainable AI - Controlable Image Generation

Introduction

This project investigates semantic disentanglement techniques to analyse the intermediate latent spaces in generative AI models, such as StyleGAN and Diffusion models. By isolating and utilising these disentangled features, we aim to develop prognostic models for lung diseases that are more interpretable for clinicians and researchers. This approach will facilitate the identification of novel biomarkers for lung disease prognosis. The detailed project in medical domain can be found in the poster.

What is XAI?

XAI is short for Explainable AI. In other words, it is the set of techniques and methods created to make understandable and interpretable to humans decision-making processes developed by AI systems. Unlike a traditional black box AI model that offers little insight into how its conclusions are drawn, XAI provides clear and concise explanations, which help a user understand why an AI-made decision was chosen.

The Black Box Problem in AI

AI systems, and especially those involving deep learning, are mostly black boxes. Thus, even when they make excellent predictions, people cannot explain what the reasonings are. Some of the associated problems are discussed below:

Lack of Trust: It would be very challenging for the stakeholders to put faith in an AI system where people cannot know why a particular decision was taken. Regulatory Obstacles: Many sectors are regulated to be transparent in the decision-making process. Ethical Concerns: Without understanding the logic behind AI decisions, businesses risk perpetuating biases and making unethical decisions.

The black box nature of AI is particularly problematic in high-stakes industries such as finance, healthcare, and law, where decisions can significantly impact people’s lives.

Letent Space in Generative AI

Result

📄 Full Poster

Further Work

Lung desease localisation and controllable generation