Zeiss Projects

Past Projects:

Taking humans out of the loop: Intelligent Microscope Systems

For an imaging system with multiple parameters/settings (e.g. laser power, focus, contrast, exposure
time etc.), it is not trivial to select a combination of parameters that generates a high quality image for a
given sample. Experienced human operators manage to obtain a good parameter set after a few trialand error iterations, but such a manual process is tedious. Moreover, the image quality achieved depends
highly on the competence of the user. While Bayesian optimization (BO) has been successful at tackling problems of automatic parameter selection, to the best of our knowledge it has not yet been thoroughly investigated for a problem where the objective function f(x) relates to an image acquisition hardware. To formulate this mathematically as a BO problem, the first step is to develop an appropriate image evaluation function that assesses the quality of an acquired image of a sample. For doing this, the BO agent does not have of course access to the ideal image of that specific sample, but it does have access to a pool of high-quality images of versatile samples from where it can derive a meaningful objective for BO. This critical step will be addressed in the first work package of the project. To navigate the parameter search space efficiently and reach near-optimal values with a minimum number of steps, the BO agent needs to incorporate information from a training process. For instance, domain knowledge should be incorporated as priors in the underling model. This will be investigated in work package 2. Moreover, to reach human-like performance in the parameter tuning process, the BO agent may not only consider the image quality metric but must also exploit information given by the image itself. e.g. reduce the laser power if the image is too bright. These add-ons to the baseline BO solution will be investigated at work packages 3 where context and gradients will be investigated from a multi output perspective.

Machine Learning Assisted Optical Inspection in Additive Manufacturing via Domain Adaptation

A promising technology to ensure high quality of printed parts in additive manufacturing is the combination of optical inspection and machine learning (ML). This project aims to advance the current capabilities of ML in additive manufacturing by reducing the need for data re-annotation any time the printing task is changed. Instead, training data will be created artificially by converting previously collected annotated data from different printing tasks to the new appearance of the task at hand.
We contend that this problem falls under the umbrella of domain adaptation (DA). The state-of-the-art DA approaches rely on the ability to generate data conditioned on synthetic or other relevant data sources that we have access. The generation process heavily relies on the generative adversarial networks (GANs) that have recently risen to stardom due to their expressivity and ease for generation on runtime. To this end, our project seeks will study not only the general GAN training problems, but also to attack the GAN training in the specific context of DA. In this setting, we will identify assumptions that help with the convergence of Robbins-Monro-type algorithms in GAN training or develop newer, scalable minimax optimization algorithms that can exploit the specific, newer structures that are inherent in the domain adaptation setting, including specific settings that could boost performance in downstream tasks.

Robust  Verification and Uncertainty Estimation with Deep  Neural Networks.