Large Language Models (LLMs) can perform a wide range of tasks, but their performance can be negatively impacted when there's a switch in tasks. This study is the first to formalize the study of such vulnerabilities, revealing that both very large and small LLMs can be susceptible to performance degradation from task-switches.

We proposed a multi-task learning paradigm for Concept Bottleneck Models to introduce inductive bias in concept learning. Our proposed model coop-CBM improves the downstream task accuracy over black box standard models. Using the concept orthogonal loss, we introduce orthogonality among concepts in the training of CBMs.

We propose Transparent {A}nomaly Detection {C}oncept {E}xplanations (ACE). ACE is able to provide human interpretable explanations in the form of concepts along with anomaly prediction. Our proposed model shows either higher or comparable results to black-box uninterpretable models.

We propose RelationalUNet which introduces relational feature transformation to the UNet architecture. RelationalUNet models the dynamics between visual and depth dimensions of a 3D medical image by introducing Relational Self-Attention blocks in skip connections.

We propose uncertainty based strategy to select the interventions in Concept Bottleneck Models during inference.