Artificial Intelligence for the Sciences Group
Prof. Dr. Frank Noé
We are always looking for highly motivated PhD students to work on exciting projects concerning artificial intelligence in the sciences.
Project 1: Deep learning of three-dimensional structures
Deep neural networks, and in particular representation-learning structures such as convolutional neural networks have set new standards and achieved superhuman capabilities in a number of classical machine learning problems. These structures are especially strong when the input data has an natural embedding in a physical space (such as the 2D pixel space of images) and efficient filters can be found that capture local features and patterns and can be combined to higher-order features. While inference for 1D structures (text, speech, sequences) and 2D structures (images) has made remarkable process, the inference of 3D structures that are important for biology is still in its infancy. We propose projects to design and employ deep neural networks for 3D reconstruction problems such as protein structure prediction from sequences (1D->3D) or from Electron micrographs (2D->3D).
In collaboration with the Structural Biochemistry group of Prof. Dr. Markus Wahl
Project 2: Learning spatiotemporal patterns from time-series data
High-dimensional time series are ubiquitous in biology (e.g. molecular dynamics simulations, single particle tracking microscopy of proteins moving in a cell, electroencephalography of brain activity) and other fields (e.g. climate and weather data, finance, sports). Learning the essential spatiotemporal patterns from such data can (i) teach us to understand the nature of the dynamical systems (e.g. which neurons signal concurrently in certain thought processes, which atoms in a biomolecule move in a dynamical rearrangement, which soccer team formation and motions are most promising for scoring or preventing goals), (ii) allow us to compress data with high-dimensions and may time steps efficiently without loosing relevant information (e.g. consider the famous one-pixel camera developed by Baraniuk and co-workers at Rice), (iii) allow us to predict the future evolution from past time series (obviously relevant for weather and climate prediction, but also important to early detect epileptice seizures and other critical health problems from biosensor data). Recent theoretical works and computational methods have taught us a deep understanding of the structure of such learning problems, in particule time-lagged independent component analysis (TICA, Molgedey and Schuster, 1994), Dynamic mode decomposition (DMD, Schmid and Sesterhenn, 2008), and the Variational approach for Markov processes (VAMP, Noé and Nüske, 2013). Projects around this theme will involve method and software development including kernel methods, sparse compression and deep autoencoders.
Project 3: Deep learning for super-resolution microscopy
Super-resolution microscopy is a family of techniques that can achieve higher resolutions than the diffraction limit of optical light microscopy (250 nm) These techniques are at the forefront of generation of new insights in biology and have been awarded with the Nobel prize in Chemistry in 2014. This is achieved by either of two tricks or their combination: (1) By stochastically switching fluorescence labels off an on such that on average a sparse signal is obtained with typical distances between active labels greater than the diffraction limit, such that the positions of the fluorescence labels can be determined by fitting the mean of the distribution of localizations. In this approach (STORM, PALM), super-resolution is essentially a data analysis problem. (2) By exciting the sample with nonstandard laser profiles or using complex excitation patterns (STED, LLS), one essentially uses higher modes of the signal to probe finer structures.
In the last few years, deep learning methods have made significant progress in essentially all areas of science and engineering. Deep learning has recently been demonstrated to be able to help achieve super-resolution from diffraction limited microscopy, to help in localizing fluorescence labels in STORM and PALM, and to imrove the statistical efficiency of super-resolution techniques. As in this field, higher resolution almost d irectly translates to new biological insights, we want to develop novel deep learning methods that can enhance the effective resolution (or increase the data efficiency at fixed resolution) for state-of-the-art imaging techniques, such as lattice light sheet microscopy.
In collaboration with the Membrane Biochemistry group of Prof. Dr. Helge Ewers
For more information, have a look at the website of the Artificial Intelligence for the Sciences group at FU Berlin