Ioannis Sarridis presents Spurious Correlation Biases: Challenges and Mitigation Approaches.
On 2026-07-01 - 2026-07-01 11:00:00 at G205, Karlovo náměstí 13, Praha 2
Spurious correlations occur when certain attributes in the training data
strongly co-occur with target labels without reflecting a true causal
relationship. Models trained on such data tend to rely on these shortcuts
instead of task-relevant features, leading to poor generalization under
distribution shifts. In this presentation, we will examine this problem in the
context of computer vision, where such correlations often arise from
backgrounds, textures, co-occurring objects, demographic attributes, or other
visual attributes that are difficult to isolate. We will discuss the challenges
of identifying and mitigating spurious correlations in image data, and present a
set of methodologies that progressively relax the assumptions required to
address them.
strongly co-occur with target labels without reflecting a true causal
relationship. Models trained on such data tend to rely on these shortcuts
instead of task-relevant features, leading to poor generalization under
distribution shifts. In this presentation, we will examine this problem in the
context of computer vision, where such correlations often arise from
backgrounds, textures, co-occurring objects, demographic attributes, or other
visual attributes that are difficult to isolate. We will discuss the challenges
of identifying and mitigating spurious correlations in image data, and present a
set of methodologies that progressively relax the assumptions required to
address them.