We address the problem of optimizing the parametrization of a super-resolution STED microscope under variable biological conditions. The traditional approach would be to perform a grid search in the space of possible configuration parameters. Such a uniform coverage of the space causes the acquisition of several unusable images due to their poor parameterization. In this application, we aim to find the optimal configuration while maximizing the quality of the images taken during the search. We tackle this application under the bandits setting.

In many situations, it is not sufficient to limit oneself to optimizing the quality of the images: different potentially conflicting objectives must be taken into account. For example, when an image is acquired, an attempt to increase the quality of an image may lead to degradation of structures and fluorescent markers, the effects of which can be observed in the same image. It is therefore necessary to trade off between the quality of the acquired image and the preservation of the sample. We show how multi-objective optimization can be directed by introducing an expert user into the learning loop.

Le séminaire sera présenté à 11h30 au Pouliot-1120.