Prof. dr. Michael S. Lew

Professor, Hoogleraar Deep Learning
Head, Deep Learning Research Group
Head, Machine Learning Cluster (30+ members)
co-Director, LIACS Media Lab
co-Head, Computational Imaging and Deep Learning Research Group
Computer Science Department (LIACS)
Leiden University
lewmsk@gmail.com

My Positions, Education and Research

Full Professor, Hoogleraar Deep Learning at LIACS, Leiden University

Head, Machine Learning Cluster (30+ members)

Co-head of the LIACS Computer Systems, Imagery and Media Research cluster (70+ members), which is one of the two main research divisions at the computer science department at Leiden University.

PhD, University of Illinois Urbana-Champaign (UIUC)

Research - short version

My research lies in the field of artificial intelligence with a primary focus on the area of deep learning. I develop new deep learning algorithms toward understanding large databases and libraries of information by learning synergistic methods using multiple modalities within multimedia including but not limited to vision, audio and text. The goal is to bring to light all of the information in the world and at the same time automatically be able to detect misinformation (e.g. fake news).

Research - longer version

My research is toward developing artificial intelligence (AI) and currently focuses on the paradigm of deep learning.

Artificial Intelligence is the most promising paradigm to help society. Major breakthroughs are expected in the near future in a numerous important areas of society (e.g. medicine, automatic self-driving vehicles, fake news detection, robotics, etc.) but only with access to very large datasets. In many areas, the datasets exist but scientists do not have access. An example I am often asked by medical researchers: what it would take to have expert level classifiers for well known diseases and cancers. My reply is usually that one needs access to a very large dataset of examples. These datasets exist already at many research hospitals around the world. However, the datasets can not be used for training the deep classifiers due to privacy reasons. In my opinion, while it is important to develop AI to detect diseases and cancers, it is even more important to protect human privacy rights. Currently, there is no easy answer for the dataset access issue and each country is wrestling with the tradeoff of privacy vs scientific advances in their own way.

My current research is exploring how to both find information and also how to detect false information. Both are necessary to a modern information retrieval system. Because humans want to express their search queries using high level human concepts, the grand challenge in multimedia retrieval has been "bridging" the semantic gap which is the gap between the high level concepts of humans and the low level features from multimedia. For a specific example, the problem of image classification is especially important because it can bridge the semantic gap by taking an input image and outputting human level concepts. If we can solve the semantic gap, then it would make all of the collected art/heritage, scientific, WWW and personal images searchable and accessible. Furthermore, there is major interest in explanation based modeling which seeks to give better understanding of the underlying reasons and causes involved in deep network decisions.

Misinformation is one of the key 21st century problems and manifests in many different ways. "Fake News" is perhaps the most well known type and has affected areas from medicine (dangerous treatments for COVID) to politics (unfounded claims of election fraud). Misinformation also occurs in more subtle ways such as intentionally misrepresenting someone's viewpoints such as taking a small part of a discussion and presenting it as the main viewpoint. The context is important. Other modern ways in which misinformation is happening is with misrepresentation of authorship (see ChatGPT) and plagiarism. Part of my current research examines methods of detecting fake information using deep learning.