{"value":"

Amsterdam Data Science in collaboration with University of Amsterdam, Vrije Universiteit Amsterdam and Prosus are hosting a webinar on Universal Representation.

A challenge for most organisations is finding the best way to combine data from different sources to facilitate and streamline machine learning advances. This complex subject sweeps multiple domains: infrastructure, methods and algorithms, company organization, tools and libraries, among others.

During this Meetup, the goals is to share and learn about an important subtopic: how we can represent complex objects for the purpose of machine learning, including methods, techniques, advantages and disadvantages, lessons learnt, and what the future holds.

The closest term for what we want to explore is universal representations. Universal representation assumes a slightly different meaning depending on what academic sub-field you are looking at. By universal representations, we are talking about embeddings of complex objects that are useful for downstream tasks.

Talk #1 by Stevan Rudinac

I am an Associate Professor of Artificial Intelligence for Business at the University of Amsterdam Business School and a guest researcher at the Informatics Institute of the UvA. In my research I aim at enabling large-scale multimedia analytics based on the relevance criteria defined at a higher semantic level, by jointly analysing visual content and the heterogeneous information associated with it, ranging from text, automatically generated metadata and open data statistics to information about users and their social network. What fascinates me is the potential of artificial intelligence in addressing important societal challenges, such as liveability and security.

Multimodal Representation Learning
Abstract follows.

Talk #2 by Michael Cochez

Dr. Michael Cochez is an assistant professor in the Knowledge Representation and Reasoning group at the Vrije Universiteit Amsterdam and manager of the Discovery Lab (an ICAI lab in collaboration with Elsevier and the University of Amsterdam). He works on bridging the gap between machine learning and knowledge graphs. His research interests include embedding of knowledge graphs for downstream machine learning tasks, dealing with missing information in graphs (link prediction, approximate graph query answering) and applications such as question answering and recommendations. dr. Cochez obtained his BSc. degree from the University of Antwerp, Belgium, his MSc. and Phd from the University of jyväskylä, Finland, and was a postdoc at Fraunhofer FIT, Germany. More info https://www.cochez.nl/

Bridging the Gap between Knowledge and Learning

Machine learning approaches have been applied successfully in many fields, also in knowledge representation. For example, graph embeddings techniques have been taken up by the community as a tool to solve various tasks. Also other models which connect the knowledge representation, where information is typically represented in the form of a graph, with the machine learning world exist. In this talk, I will first give a general overview of the topics in the field of representation learning on knowledge graphs and then present our recent work on inductive representation learning and approximate graph query answering.

Talk #3 by Riccardo Campari

Riccardo Campari is a Data Scientist at Prosus, the largest consumer internet company in Europe. He has a physics background (PhD from U Parma and postdoc studies at MIT's Senseable City Lab) and has worked in trading and finance prior to joining Prosus. He works across multiple companies in Prosus’ portfolio, with a focus on predictive modeling, technological experimentation, and automation.

Enabling Rapid AI in Complex Environments

Consumer internet companies deal with enormous quantities of data: images, sound, video, text, as well as chats, users' actions and interactions to name the most common. Efforts to combine these data sources with machine learning to improve a company's services meet with plenty of challenges; in data-rich organizations, one of the main bottlenecks is the high level of expertise needed to find, sieve, and manipulate relevant data sources. We explore here a few ways to systematically generate universal representations of arbitrary, complex entities; by minimising the need for data-specific and topic-specific knowledge, universal representations enable lightning-fast development cycles of new ML models.

Universal Representation

Speakers

The event is over

Amsterdam Data Science