Hey there!

In this project, I conducted a critical analysis of the methodology proposed by Ribeiro et al. for generating high-fidelity counterfactual images using Structural Causal Models (SCMs). The research focused on evaluating the model's effectiveness in producing realistic and causally consistent counterfactuals applied to both the MNIST and Chest X-ray datasets. I assessed the model's performance and explored its implications for causal generation. Additionally, I delivered a presentation outlining the motivation and background of causal generation and SCMs, detailing the methodology of the original paper, and sharing my findings from using their implementation.

Abstract: There is a need for the controllable generation of medical images in order to monitor chronic diseases such as Multiple Sclerosis (MS). In order to develop such models, highly correlated attributes need to be sufficiently separable so that the model can learn how each independently effects the generated image. Learning a disentangled representation of attributes would thus allow for specific control in the generation process. In this work, the latent spaces of the StyleGAN are studied to understand how disentangled they are. Disentanglement is both qualitatively and quantitatively assessed for the StyleGAN and its variations, with metrics formulated from models specifically trained on MS data. Although the resulting generation still requires optimization, this is a first step in creating a conditional model for controllable generation in medical imaging with vast applications such as trajectory modeling and treatment effect.

Abstract: Although exact inference in Probabilistic Graphical Models (PGMs) is possible on relatively small graphs and in tree-like graphs by using message passing algorithms such as Belief Propagation (BP), performing these tasks on graphs of arbitrary size and connectivity is computationally expensive and often intractable. Methods like BP can be used for approximate inference but generally struggle in loopy graphs or graphs with far-reaching node dependencies. Although used in different applications, there is a striking parallel between the fundamental concept of message passing in PGMs and Graph Neural Networks (GNN). This naturally gives way to using GNNs for approximate inference in PGMs by mapping nodes in PGMs to nodes in GNNs. In this work, we compare BP with GNN methods both in tree and non tree graph structures. To extrapolate the need for importance weighing in graph structures, we explore the use of soft attention in GNNs. Expanding on the works of Yoon et al. (2019), our work demonstrates that using this mechanism in the GNN outperforms its variants and BP on various graphs. Our code is publically available in the link below.