Towards trustworthy computer vision models

Speaker: Tatiana Tommasi, Politecnico di Torino

Deep Learning models are becoming ubiquitous in our daily lives, we often use AI solutions even without realizing it. In this talk I will focus on computer vision approaches and discuss relevant aspects that can make them trustable. Indeed, they should provide high accuracy even in open-world conditions, facing novel categories and novel visual domains. Moreover, they should guarantee fairness, avoiding to rely on data that contain stereotypes and bias on sensitive attributes. Finally, they should be robust to hardware and software faults. I'll present some examples of these cases, discussing possible solutions and future directions.

Zoom