-----

 

In recent years, deep neural networks have rapidly approached human visual capabilities through end-to-end training on semantic tasks such as object recognition and vision-language alignment. However, research at the intersection of deep learning and psychophysics has also revealed striking differences to human visual perception, such as a lack of robustness and non-human-like errors. Tangemann's research aims to improve the alignment of human and machine visual perception with a focus on mid-level vision, and in particular perceptual organization: How do humans and machines organize the signals from millions of photoreceptors and pixels, respectively, into a coherent, object-centered scene representation? Tangemann will present recent studies centered around the Gestalt principle of "common fate" and discuss how the close link between motion and objectness can be modeled using DNNs. These studies not only reveal further differences between humans and DNNs in motion perception, but also offer a perspective on how insights from vision science can help improve the alignment of human and machine vision.

-----