The goal of lifelong machine learning is to develop techniques that continuously and autonomously learn from data, potentially for years or decades. During this time, the system should autonomously improve its performance by extracting and preserving information between different learning tasks, similar to how a natural system learns more and more complex tasks over time. In my talk, I will highlight recent work from our research group in two directions: theoretical guarantees for lifelong learning and applications to computer vision problems.
The communication presentations (keynote talks, facilitated break-out sessions, posters) will be continuously in May and June.
Christoph Lampert received the PhD degree in mathematics from the University of Bonn in 2003. In 2010 he joined the Institute of Science and Technology Austria (IST Austria) first as an Assistant Professor and since 2015 as a Professor. His research on computer vision and machine learning has won several international and national awards, including best paper prizes at CVPR and ECCV in 2008. In 2012 he was awarded an ERC Starting Grant by the European Research Council. He is an Editor of the International Journal of Computer Vision (IJCV), Action Editor of the Journal for Machine Learning Research (JMLR), and Associate Editor in Chief of the IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI).
|CHIST-ERA Conference 2016 - Christoph Lampert.pdf||533.42 KB|
The poster presents the possibility of direct comparison of medical text content by using unstructured representation of document information in frequency matrix of terms. Dimensionality reduction is performed using Latent Semantic Indexing method. Two common metrics are used: Cosine distance and Jaccard metric. Cosine measure shows a lower sensitivity for finding similar documents. The analysis was performed on set of 400 cases of description of abdominal radiological diagnostic images.
We focus on a multiagent systems (MAS) scenario of human and artificial agents capable of learning. We assume artificial agents are capable of case-based learning and indicative learning, from which they are capable of explain and justify the hypotheses they have learnt. Given the social context of multiple individual agents, disagreements can be aggressed by arguments pro and against specific learnt hypothesis. Since humans are adept at conversational argumentation, the approach encompasses scenarios with MAS in which the agents can be both human and artificial. We have developed AMAIL, a first platform for artificial agents that integrate symbolic learning and computational argumentation. The approach covers the following human/artificial learning agents scenarios: (1) the student/apprentice scenario, (2) the multiple expertise individuals scenario, and (3) the lifelong human supervision of learning system via dialogue (in which the learning system has to be convinced to change hypothesis via conversation instead of blindly trusting an infallible oracle).
|CHIST-ERA Conference 2016 - E. Plaza.pdf||439.48 KB|
The diagnostic approach in Alzheimer's disease is still based on anamnesis and neuropsycological assessment besides laboratory tests and brain scanning addressed to rule out other conditions that can mimic the clinical picture. Biomarkers, which have provided a tool for supporting the diagnosis at the time of improving the understanding of the neuropathological processes, are not being widely used in clinical practice yet. Modular knowledge models providing measurable accuracy of the diagnosis based on the up-to-date knowledge are of main interest for a rational use of biomarkers in the clinical setting. This challenge is faced merging techniques such as automated learned ontologies, active learning and on-line learning.
|CHIST-ERA Conference 2016 - J. Villar.pdf||493.2 KB|
Justus Piater is a professor of computer science at the University of Innsbruck, Austria, where he leads the Intelligent and Interactive Systems group. He holds a M.Sc. degree from the University of Magdeburg, Germany, and M.Sc. and Ph.D. degrees from the University of Massachusetts Amherst, USA, all in computer science. Before joining the University of Innsbruck in 2010, he was a visiting researcher at the Max Planck Institute for Biological Cybernetics in Tübingen, Germany, a professor of computer science at the University of Liège, Belgium, and a Marie-Curie research fellow at GRAVIR-IMAG, INRIA Rhône-Alpes, France. His research interests focus on visual perception, learning and inference in sensorimotor systems and other dynamic and interactive scenarios, and include applications in autonomous robotics and video analysis. He has published more than 150 papers in international journals and conferences, several of which have received best-paper awards, and currently serves as Associate Editor of the IEEE Transactions on Robotics.
Scaling up the ability of robots to interact with uncontrolled environments in sophisticated ways requires incremental learning capabilities that allow the robot to exploit already-learned concepts to facilitate future learning. I present two examples of such systems from my lab, and discuss challenges on the way towards scalable robot learning.
Analytic provenance captures the context and process of the sensemaking/decision-making, including the data explored, findings, hypothesis, reasoning process, and their relationships. In our previous study, we have shown that the visualisation of the analytic provenance information can be used to support sensemaking, for purpose such as exploration, collaboration, and reporting. We believe it has a role in tracking and dealing with uncertainty as well, by show the source the data (uncertainty), the error introduced by analysis algorithm, and any human bias during the reasoning process.
Matthias Harders received his PhD degree and completed his habilitation at ETH Zurich, Switzerland, in 2003 and 2007, respectively. There he founded and led the Virtual Reality in Medicine research group. In 2014 he became a full professor at the University of Innsbruck, where he now leads research on interactive graphics and simulation. He is a co-founder of the IEEE Technical Committee on Haptics, of the EuroHaptics Society, and the IEEE Transactions on Haptics. In 2008 he also co-founded the start-up company VirtaMed, which builds surgical training systems. His current research focuses on physically- based simulation, computer haptics, and virtual/augmented reality. The main application area is the medical domain, also including aspects of human- computer interaction and multi-modal data visualization.
Despite trends towards increased automation, human users are still frequently central in interpretation, analysis, and decision making based on data. A key challenge in the context of Big Data is the design of intuitive, interactive, and user-friendly interfaces, that allow natural and effective working with data. In this regard, displaying information not only via the visual channel, but also others, such as audio or haptics, has potential to reduce the cognitive load. In this talk examples of existing systems employing haptic data visualization will be given, and possible future research directions will be outlined.
In this topic we propose to further develop the most promising machine learning algorithms for image processing by bringing the user in the loop. The traditional approach is to create a training dataset to train the algorithm (using e.g. deep learning, low level features, Haar cascades or similar) to verify the algorithm using the testing dataset and to finally release the algorithm to the industrial application. We believe that this approach differs from the effective way humans learn – that is learning by doing and evaluating. We propose that new algorithms should be developed, that are capable to digest the feedback from the final user in order to improve. Practical applications may be considered in all areas where the algorithm outcome is verified by human, such as security applications, quality control or multimedia analysis.
|CHIST-ERA Conference 2016 - M. Grega.pdf||3.08 MB|
- 1 of 2
- next ›