Vancouver, 20-27/02/2024.
The AAAI conference promotes research in Artificial Intelligence and fosters scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines.
At this year’s event, AAAI hosted a special track on Safe, Robust and Responsible Artificial Intelligence (SRRAI). Scientists from Fraunhofer HHI, partners of the iToBoS project, presented their paper titled “From Hope to Safety: Unlearning Biases of Deep Models via Gradient Penalization in Latent Space” in an oral session. In this work, we introduce a novel bias mitigation approach that penalizes the model if it uses features, measured via the latent gradient, pointing into a concept direction representing the bias.
The paper resulted from research conducted throughout the iToBoS project. Several confounders occur in datasets used to train neural networks for Melanoma detection, such as rulers, band-aids, hair, or skin markers. Our new bias mitigation method can reduce the model’s sensitivity towards these data artifacts.
Find out more at https://aaai.org/conference/aaai/aaai-24.