Self-supervised learning for computer vision

In this brief talk I will give an overview of my research on self-supervised learning (SSL) for computer vision. We will dive into how clustering can be combined with representation learning using a single loss and how semantic the resulting clusters are (ICLR'20). Next, we take a look at how clustering can be generalized to audio-visual datasets (NeurIPS'20) and how many SSL pretext tasks can be expressed using a single contrastive framework that only distinguishes between transformations to be invariant, or variant to (ICCV'21). Finally we talk about two works that tackle ethics & privacy in AI: first, an analysis of language models "out-of-the-box" with regards to their intersectional biases in the context of occupations (NeurIPS'21) and second, a novel large-scale dataset without any pictures of humans as a more privacy-secure replacement for ImageNet for the purpose of developing self-supervised pretraining methods (NeurIPS-Data'21).

You can find the papers mentioned here: https://yukimasano.github.io/

Short Bio:
Yuki M. Asano is an assistant professor for computer vision and machine learning at the University of Amsterdam and Lab Manager for the QUVA Lab, working with Cees Snoek, Max Welling and Efstratios Gavves. Prior to this, he completed his PhD student at the Visual Geometry Group (VGG) at the University of Oxford where he was advised by Andrea Vedaldi. His main research interests lie in self-supervised and multi-modal learning and more recently bias and privacy issues in AI. Prior to the PhD, he studied physics at the University of Munich (LMU) and Economics in Hagen as well as a MSc in Mathematical Modelling and Scientific Computing at the Mathematical Institute in Oxford.

https://uva-live.zoom.us/j/85418373321
Meeting ID: 854 1837 3321