Emily Denton is a Research Scientist in Google's Research and Machine Intelligence group where she works on developing tools and techniques to promote fair, inclusive, respectful and ethical AI. She is particularly interested in detecting and mitigating harmful bias in computer vision systems and understanding the social impact of machine learning technology. Prior to joining Google, Emily received her PhD in machine learning from the Courant Institute of Mathematical Sciences at New York University in 2018. Her research focused on unsupervised learning and generative modeling of images and videos. Emily has been awarded a Google Fellowship in Machine Learning and enjoyed support by the Natural Sciences and Engineering Research Council of Canada (NSERC).

Deep learning has revolutionized the field of computer vision in recent years. Large scale automated image understanding systems impact the daily lives of people around the world and hold the potential for great positive impact. However, many historical biases and patterns of discrimination frequently embed themselves into the data and models being deployed. In this talk we track how harmful and unintended bias can enter computer vision systems through data collection, data annotation and system design. We will emphasize the importance of considering the social context within which computer vision models are built and deployed. We also discuss the ethical considerations of the research and development of certain classes of computer vision technology. Finally, we will dig into recent research on detecting and mitigating unfair bias and perpetuation of historical patterns of discrimination in computer vision technology.

Jussi Karlgren is an adjunct professor of language technology at KTH and a founding partner of the text analysis company Gavagai where he spends most of his time. He holds a PhD in computational linguistics, a Ph Lic in Computer and Systems Sciences, both from Stockholm University. He also holds the title of docent (adjoint professor) of language technology at Helsinki University. He has worked with research and development in information access-related language technology at IBM Nordic laboratories (1987-88), at the Swedish Institute of Computer Science (SICS) (1990-2010), at Xerox PARC (1991), at New York University (1995-96), at Yahoo! Research in Barcelona (2007-08), and at the department of linguistics at Stanford University (2017-18).

His research interests are modelling genre and stylistics, studying the negotiation of meaning in human communication, and on evaluating and validating information systems. He has previously worked on projects on natural language interfaces to databases, speech-based route guidance systems, plan inference, human-computer dialogue in virtual realities, on automatic speech-to-speech translation, on modelling and raising awareness of energy usage in homes and workplaces, and on citizen observatories.

High-dimensional distributed semantic spaces have proven useful and effective for aggregating and processing visual and lexical information for many tasks related to heterogenous data generated by human behaviour or relevant for human information processing. Models to process such high-dimensional spaces have proliferated in recent years with impressive results on quite various tasks. The representation in such models is usually not designed to be transparent or humanly manipulable.

In general, a representation used in research should hold to some basic qualities. It should have descriptive and explanatory power; be practical and convenient for further application; allow generalisations to be made and analogies to be inferred; be reasonably true to human performance, providing defaults to smooth over situations where data are lacking and constraints where the decision space is too broad, perform seamlessly and incrementally online in face of novel data, allowing new features and new analyses to be incorporated without recompiling previous understanding. It should also-in view of future application to downstream tasks-be computationally habitable.

Human language has a large and varying number of features of various regularity, incorporating both lexical items and constructions. The field of linguistics provides a large body of research to understand such regularities. Yet the models used to process human language disregard those regularities, starting from the general principle that they should be discovered rather than given and that learning should be in the form an end-to-end classifier from raw data to relevant categories. This principle is intuitively appealing, in view of the specificity and avowed situation- and task-independence of linguistic rules, but the tools built end up being black boxes and do not guarantee explanatory generality of the results.

This talk will show how high-dimensional semantic spaces can be editorially assembled from features designed to test hypotheses, and how models previously used to handle word semantics can be straightforwardly extended to handle more complex linguistic items. Such models can then represent broad ranges of linguistic features in a common integral framework which is suitable as a bridge between symbolic and continuous representations, as an encoding scheme for symbolic information.

Geert Litjens studied Biomedical Engineering at Eindhoven University of Technology. Subsequently, he completed his PhD in the Diagnostic Image Analysis Group. He worked with Henkjan Huisman on Computer-aided detection of prostate cancer. He spent 2015 as a postdoctoral researcher at the National Center for Tumor Diseases in Heidelberg, Germany on an Alexander von Humboldt Society Postdoctoral Fellowship. In 2016 he was awarded a Bas Mulder Award by the Dutch Cancer Society with which he returned to the Diagnostic Image Analysis Group. In 2017 he was awarded a Veni grant by the Netherlands Organisation for Scientific Research. He is currently an Assistant Professor in Computational Pathology at the Department of Pathology.

Machine learning, especially in the form of deep learning, has pervaded large parts of society. Even in healthcare, an area notorious for the slow uptake of new technology, has seen its first FDA-approved machine learning algorithms enter the market. Designing algorithms for healthcare is challenging, not just due to the fact that a wrong prediction can literally cost lives. Doctors also need to want to use algorithms, which means they need to solve meaningful tasks, be reliable and should be implemented efficiently in their workflow. In this presentation, I will discuss several practical examples of such algorithms.

A second challenge is reliable evaluation of algorithms for healthcare in a robust, reproducible manner. With the CAMELYON challenge we tried to establish a first benchmark for reliable, large scale evaluation of computational pathology algorithms. I will discuss the lessons learned from organizing challenges and the setup for potential future challenges.

Wiro Niessen is full professor in Biomedical Image Analysis at Erasmus MC, Rotterdam where he leads the Biomedical Imaging Group Rotterdam and at Delft University of Technology.

His interest is in the development, validation and implementation of machine learning and quantitative image analysis methods in clinical practice and biomedical research, in linking imaging and genetics data, radio(geno)mics, and image guided interventions. Focus areas are improved diagnosis and prognosis of neurodegenerative and cardiovascular disease, and treatment guidance in oncology, by developing and using advanced medical image analysis and machine learning methods. He has published over 300 journal articles in these areas, and has a H-index of 75 (Google Scholar).

He is fellow and president of the Medical Image Computing and Computer Assisted Interventions Society (MICCAI), director of the Biomedical Image Analysis Platform of the European Institute of Biomedical Imaging Research (EIBIR), member of the International Society of Strategic Studies in Radiology (IS3R), board member of the Technological Branch of the Netherlands Organization of Scientific Research (NWO-TTW) and Chief Digital Officer of Health RI, a national research infrastructure for prevention, personalized treatment, and health.

In 2012 he founded the spin-off company Quantib, which is an AI company in medical imaging that develops methods to automatically extract relevant information from medical imaging data. He is currently the Scientific Lead at Quantib. In 2015 he received the Simon Stevin Master award, the largest prize in the Netherlands in the field of Applied Sciences. In 2017 he was elected to the Royal Netherlands Academy of Arts and Sciences.

The combination of big data and artificial intelligence are dramatically increasing the possibilities for prevention, cure and care, and are changing the landscape of the healthcare system. Medical imaging will play a central role in this revolution, and is one of the first domains in medicine in which these techniques are being employed. Among the AI techniques, machine learning, and especially deep learning is a disruptive technology in this field, impacting medical image acquisition, reconstruction, analysis, and image-based diagnosis and prognosis.

In this presentation I will show the opportunities and challenges of big data analytics with AI techniques in medical imaging, also in combination with the analysis of genetic and clinical data. Both conventional machine learning techniques, such as radiomics for tumor characterization, and deep learning techniques for studying brain ageing and prognosis in dementia, will be addressed. Also the concept of deep imaging, a full integration of medical imaging and machine learning, will be discussed. Finally, I will address the challenges of how to successfully integrate these technologies in daily clinical workflow.

Josephine is a lecturer in computer vision within the School of Computer Science and Communication (CSC) at KTH, the Royal Institute of Technology, Stockholm, Sweden. She belongs to the CVAP group and on a even more fine-grained classification I'm part of the Computer Vision Group within CVAP. Her research efforts are devoted to the field of computer vision. In the past she has worked on the problems of visual tracking, human/object pose recognition and multi-target tracking. However, for the last couple of years she has been focusing on deep learning for solving computer vision problems.

Josepine obtained her D.Phil from Oxford University, Dept. Of Engineering Sciences in 2001 under the supervision of Prof. Andrew Blake. Prior to that she did a BA in Mathematics at Trinity College Dublin, Ireland. After her D.Phil studies she made her way to Sweden to work with Stefan Carlsson at KTH for a post-doc.

In 2014 the simplest and most effective way to solve many computer vision recognition tasks, given a limited set of labelled training data, was: 1) Take a large modern ConvNet trained on ImageNet 2) Adapt the final layer to the task at hand and then (optionally) 3) Fine-tune your new network with your given task data. And since 2014 this approach has been widely adopted both within academia and industry for a broad range of vision applications. In fact transfer learning of this type is also increasing popular for NLP and speech. But today has this basic formula been updated and improved upon? In this talk I will give an overview of transfer learning, leveraging deep learning, and its development in the last 5 years.