Masterclass on Machine Learning for Inverse Problems

I took one course from the semester program called Masterclass on Machine Learning for Inverse Problems. Right now I live in Paris but I came back in May to participate in the course.

It was a two-day event with theoretical lessons in the mornings and practical sessions in the afternoon to run scripts in class that were prepared by lab members of the lecturers. We were encouraged to test methods/algorithms developed by the lecturers and their colleagues (included in the scripts) on our own data (real/experimental data if possible, or simply any data of our interest) and edit the scripts as we please to better understand and explore the methods/algorithms while hosts of the practical sessions provided help to debug, explain, propose suggestions, and discuss.

Right now, I am doing my PhD at the physical chemistry lab of Institut Curie. My project is focused on studying physicochemical properties of membrane proteins involved in contact between cellular organelle membranes by means of in vitro reconstitution and cryo-electron microscopy (cryo-EM). Since the materials we work with are susceptible to irradiation damage, only low electron dose could be used for imaging with cryo-EM, yielding inherently noising imaging data. Extensive computation and manual intervention are required to extract and boost meaningful signal from such data. Naturally, a major part of my work is dedicated to the development of an innovative image processing workflow by integrating classical and machine learning based algorithms to automatically and efficiently sample small membrane proteins for subtomogram averaging, which is a procedure performed to boost signal-to-noise ratio in cryo-electron tomography (cryo-ET) data and compute 3D models from the signal. As many challenges I face in image processing are inverse problems (e.g., deconvolution, denoising, & reconstruction), and I would like to explore more machine learning based solutions for them, I found this course interesting.

I already learnt about machine learning for image processing and worked with convolutional neural networks (CNN), but during the course I was introduced to more advanced concepts and up-to-date methods (e.g., the Feynman-Kac theorem for stochastic process simulation, and the SafeBayes algorithm for adapting Bayesian learning rate to data). More specifically, I learnt about the mixed scale dense CNN, which could be used for segmenting objects in my imaging data and has advantages over the commonly used U-net in memory usage, the number of parameters to tune and the amount of input required for training. It was also eye-opening to learn about machine learning methods used in the study of inverse problems in other fields (e.g., finance & fluid dynamics). And it was very interesting to hear that my fellow attendees are working on seemingly very different PhD projects in geology, medical imaging, and others, but we work with the same or very similar algorithms.

Yes, I would recommend this program! It was well organised; it was only two days, but we were introduced to a wide range of topics. The lectures were well-delivered because simple examples were often used to help students with understanding. It was very engaging as well because it was rather small scale, about 20 people attended. The practical sessions were very helpful for gaining understanding of the methods introduced, the application of those methods, and inspiring new ideas in our work. During breaks and at the end of each day, we also had time to socialize with students and the lecturers for discussions. I really enjoyed it!