Speaker 1: Mr. Parsa Hosseini, Ph.D. Student, Sharif University of Technology

Title: Mitigating Spurious Correlation in Images by Intervention

Abstract: Though proven to be strongly effective on in-distribution data for image classification, standard ERM training fails when faced with out-of-distribution samples. One case of such failure is when there is a high spurious correlation between some features of the data and the property of interest. Deep Feature Reweighting (DFR) was proposed to face this challenge by retraining the last layer of a model on a group-balanced subset of the data to reduce the dependence of models on spurious features and further enhance their attention to the already learned core ones. Although feature reweighting can alleviate relying on spurious features, the last layer features may not include pure core features. We propose a method based on interventions on input images to lessen the fake correlation between the spurious sections and the labels while retraining the last layer of a model. Based on our observation that models trained with ERM still highly attend to the core, causal part of images, we first generate masks for images using class activation maps. Afterward, we make two types of interventions on images by masking or combining them and retraining the last layer of the model on the augmented data. Along with its high interpretability, this method only needs data group labels for the model selection phase and has an overall better worst group accuracy compared to previous methods with the same amount of supervision on the group labels.

Speaker 2: Ms. Laura Sudupe Medinilla, Ph.D. Student, King Abdullah University of Science and Technology

Title: Spatial profiling in translational research

Abstract: Spatial transcriptomic technologies provide valuable insights into gene expression patterns within tissues spatial context. However, each tissue type and disease has its own challenges. Here we will discuss different approaches to understand tissue transcriptomics landscape and future challenges.