Visualization for understanding and developing machine learning

Machine learning is in various flavors being deployed in a wide range of application domains, including autonomous vehicles, robot navigation, interaction systems and even medicine. As with most large-scale data driven approaches it is in most practical cases, e.g. in deep learning architectures, hard to analyze and understand what is the underlying learned model, how accurate is it, and exactly how it is solved?

In this project, we will develop interactive visualization methods making both the learning process and the efficacy of the solution transparent to developers and users. Using deep learning and computer vision tasks as the initial application domain we will take a holistic approach and aim to investigate ways of visualizing aspects of the training data, network structures, and inference results jointly in the same framework.
Visualisering för förståelse och utveckling av maskininlärning - deep learning.
The visualization system will combine new and traditional data visualization methods with novel holistic approaches specific to visualization of machine learning systems and target the needs of both developers and systems architects. For developers, the visualizations will be integrated into an environment with tools for analyzing training data quality and feature variation, as well as the performance of the system.

Enabling informed choices

Intuitive visual models will enable informed choices when data is collected and the architecture is trained. Another key challenge is to analyze and quantify the domain shift when training and validation data are from different sources. For systems architects, the tool-set will act as a visual analysis and debugging tool for architecture design and systems analysis. In our prototype system, we will include an integrated development environment with interfaces to, and visual debugging tools for, machine learning libraries such as TensorFlow. New tools for high level architecture design and analysis is one of the most important challenges in the development of next generation machine learning algorithms.

Researchers

Anders Ynnerman

Professor, Head of Division

Daniel Jönsson

Associate professor, Head of Unit

Gabriel Eilertsen

Assistant professor

Jonas Unger

Professor

External partners

School of Computer Science and Engineering, NTU, Singapore

Jianmin Zheng, asjmzheng@ntu.edu.sg

Publications

Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafal K. Mantiuk, Jonas Unger (2017)

HDR image reconstruction from a single exposure using deep CNNs

ACM Transactions on Graphics , Vol.36 Continue to DOI

WASP research at MIT

Efficient data representation for visual machine learning

Image synthesis and efficient data representations for visual machine learning

Driven by the high accuracy and performance of deep learning algorithms in computer vision tasks, we are investigating the use and production of highly realistic training data-sets for computer vision applications.

Orchestrating machine learning development in complex medical imaging environments

Even though machine learning algorithms have shown promising results in many medical image analysis tasks there are still a lot of challenges that need to be solved before these advances can be used in clinical practice.

Tool for training a machine learning system to classify tissues.

Interaction with autonomous agents for medical decision support

The aim of the research is to explore the interaction between humans and computers by the design of contexts that create efficient ensembles of medical practitioners and learning machines.

visual methods to guide and focus the user’s attention

Focal/peripheral data visualization for industrial control rooms

The amount of operation and maintenance data delivered to the industrial control room operator grows manifold. In the scope of a PhD project, we are working with the problem of information overload.

Automatic pattern tracking and visualization for decision support

Monitoring autonomous systems plays a key role in understanding the systems’ functionality and in ensuring their safety. Without appropriate data reduction and visual support this is hardly possible.

WASP at MIT

Our research environment is located at the division of Media and Information Technology on Campus Norrköping. Our work draws on 20 years of research, focusing on visualization and interaction in the context of AI, autonomous systems and software.