Speaker: Vidya Prasad (Eindhoven University of Technology)
Abstract:
I will briefly present an overview of my PhD research, which focuses on developing interactive explainable AI methods for the robust development of deep image-to-image translation models. My work primarily targets tasks with high-dimensional image outputs, which are significantly more complex than the classification tasks typically addressed in the literature. These methods are particularly relevant to generative and medical imaging applications.
In this talk, I will highlight my work on diffusion models, a class of deep generative models recognized for their ability to produce diverse, high-quality samples. Diffusion models iteratively transform noise into refined images with deep learning, and understanding this temporal data evolution is crucial for interpreting the model's learning process. However, the high-dimensional and complex nature of this evolution poses significant challenges. To address these challenges, we developed a novel method that provides a holistic view of this generative data evolution in diffusion models. It preserves the iterative context of the generative process by sampling the generative space with tailored prompts and extracts relevant attributes from intermediate outputs. It introduces an evolutionary embedding algorithm that clusters semantically similar elements, organizes them by iteration, and aligns an instance's elements across iterations to enable studying the evolution of data. We propose rectilinear and radial layouts for effective exploration. We show how our method was applied to Stable Diffusion, offering valuable insights into its generative process.
Note: A preprint is available on arxiv.
Bio: Vidya Prasad is a final-year PhD candidate at Eindhoven University of Technology, advised by Prof. Anna Vilanova and Dr. Nicola Pezzotti. Her research focuses on developing explainable AI and visual analytics methods to enhance the safety and reliability of deep image-to-image translation models, particularly for generative and medical applications. With a strong background in explainable AI, deep learning, medical imaging, and software development, Vidya is passionate about creating robust and impactful AI solutions. She has contributed to several papers and patents and was a visiting researcher in Prof. Hanspeter Pfister's group at Harvard. Before her PhD, Vidya worked as a deep learning researcher at Philips and as a software developer at Amazon. She holds an MSc in computer science from the National University of Singapore.