Thesis opportunity- Missing Data and Deep Learning

What the Master Thesis is about/background to the problem to investigate

Deep learning exists in various forms, but one of the most common ways to interpret images is using convolutional neural networks. These are especially prevalent in an embedded setting where computational resources are limited.

Convolutional neural networks operate by applying multiple convolution kernels while incrementally decreasing the spatial resolution of the input. Typically, a convolution is applied as a pointwise multiplication between a kernel and the covered pixels in an image. The kernel is moved over the input so that it’s centered over all input values. In the presence of missing data, the image will have pixels under the kernel that are unknown. One can think of it as the receptive field of a filter kernel being damaged or incomplete.

What should we do about these missing values?

Several attempts have been made to tackle the presence of missing data. In many cases one can simply treat the pixels that are missing as zero, but not always.

Some examples of dealing with missing data in the convolution operation:

• MisConv – Convert the input signal to a statistical model.
• Normalized Convolutions – Reduce influence of missing data pixel by weighing.
• Sparse Convolutions – Only apply weights to valid pixels.
• Missing data aware subsampling – Missing data shrinks further down the network.

Of course, we’d also welcome other ideas to overcome the mentioned problem.

We are especially interested in approaches that work well for small and fast architectures.

The master thesis work focuses on the following/example of research questions

• How can an architecture be modified to improve performance on images with missing data?
• Can pre-processing achieve similar performance as active missing data handling?
• How much does downstream tasks improve due to active missing data handling?
• Could small transformers better tackle the incompleteness of the receptive field?

Prerequisites

To properly understand the various techniques to handle the missing data problem a solid foundation in mathematics and understanding of statistical models is beneficial. Knowledge of deep learning, network design training etc. is also of great value. Python programming experience is a requirement.

Contact

For more information about the position, contact:

Daniel Rydström, Group Manager Vision Algorithms, daniel.rydstrom@sick.se

or

Sarah Lantz, HR Business Partner, +46 739 10 99 37.

Welcome with your application 20^th of October at the latest!

SICK is a world-leading supplier of sensors and sensor solutions for industrial applications. We are 12 000 employees in 50 countries and our headquarter is located in Freiburg, Germany. SICK in Linköping is an innovation center for Machine Vision and we are 90 committed employees with a big interest in image processing and visualization. For more than 35 years, our team at SICK Linköping has successfully developed and delivered software for technically leading products within the field of 2D and 3D vision, as well as system solutions for i.e. robot guidance and quality control.

At SICK in Linköping, we are very proud of being a healthy and attractive workplace. For many years, we have been elected as one of the best workplaces in Sweden according to the survey Great Place to Work, the latest award is from 2023. We work actively to reduce our climate footprint and we are active in various ways to contribute to the society and to increase diversity at our workplace.