SEA Double Precision: Youtube Recommendations and Cross-domain Visual Search


This Friday we'll have two talks followed by drinks.
17:30 Emma Beauxis-Aussalet/Carlo de Gaetano (the Digital Society School of HvA) Binge-watching 3h of Youtube recommendations on the climate crisis: Where would it lead you?

Recommender systems greatly impact the scope of information we are exposed to. For critical topics like the climate crisis, the filter bubbles of recommendation systems can shape our opinion and perhaps our future. Such recommendation systems are becoming instruments of social engineering, in ways we might control or not. We will present a project that investigates the filter bubbles that arise from Youtube recommendations on the topic of the climate crisis. We will discuss our methodology to trace Youtube recommendations for 3 user profiles: climate activists, climato-skeptics, and split-minded users. Such profiles would be created by initiating prior searches on pre-selected contents, and by controlling the tracking cookies of search engines. We will present our strategy for analysing and visualizing our results, and will invite the public to give feedback and criticism on our approach.

Emma Beauxis-Aussalet is a Senior Track Associate of the Digital Society School of HvA. She holds a PhD from CWI & Utrecht University on statistics and visualization for assessing classification error and bias. She studied computer science, communication, and design. She is now leading projects related to AI for social good, and addressing the Sustainable Development Goals proposed by the UN.

Carlo de Gaetano is the primary researcher involved in this work. Carlo is an information designer holding a Master of Communication Design from Politecnico di Milano. Carlo's research focuses on visual content analysis, images as data, and the mapping of social issues.

17:30 William Thong (ISIS, University of Amsterdam) Cross-domain Visual Search

This talk discusses open cross-domain visual search, where categories in any target domain are retrieved based on queries from any source domain. Current works usually tackle cross-domain visual search as a domain adaptation problem. This limits the search to a closed setting, with one fixed source domain and one fixed target domain. To make the step towards an open setting where multiple visual domains are available, we introduce a simple yet effective approach. We formulate the search as one of mapping examples from every visual domain to a common semantic space, where categories are represented by hyperspherical prototypes. Cross-domain search is then performed by searching in the common space, regardless of which domains are used as source or target. Having separate mappings for every domain allows us to search in an open setting, and to incrementally add new domains over time without retraining models for existing domains. This is joint work with Pascal Mettes and Cees Snoek.

William Thong is currently a PhD student at the Intelligent Sensory Information Systems lab of the University of Amsterdam, under the supervision of Arnold Smeulders and Cees Snoek. Previously, he completed an M.Sc. in Biomedical Engineering from Polytechnique Montréal and Mila on the classification of biomedical images with deep learning.