SC²S Colloquium - April 06, 2016
|Date:||April 6, 2016|
|Time:||3:00 pm, s.t.|
Adrian Sieler: Refinement and Coarsening of Online-Offline Data Mining Methods with Sparse Grids
Due to the Cholesky decomposition and modifications we build algorithms to tackle the problem of adaptivity in a sparse grid density estimation approach. In a framework of an Offline/Online splitting we are now able to modify the underlying Cholesky factorization of the system matrix if the grid is refined or coarsened. Thus, the cost of an Offline step may be drastically reduced, since a new factorization does not need to be applied. With introducing the Cholesky factorization and related modifications, we enlarged the list of possible decomposition methods in the Offline step and provide a full theory and implementation to perform any kind of grid changes. We embedded this methods into a data-stream based density estimation and consequently into a data stream-based classifier as well. Directly after a new data batch is processed and the density declaring coefficients are obtained in the corresponding Online step, we can include the new knowledge into the underlying sparse grid(s). After the learning is done via problem adjusted coarsening and refinement, the changes are induced into the Cholesky factor of the system matrix through our introduced algorithms.
Vinod Rajendran: Efficiently approximating chemical properties via Deep Learning
The accurate and efficient prediction of molecular properties throughout chemical compound space is a critical ingredient towards rational compound design in chemical and pharmaceutical industries. Since the recent past, Machine Learning (ML) techniques applied to ab initio calculations i.e. nuclear charges and Cartesian coordinates of all atoms, have been proposed as an efficient approach for predicting quantum-mechanical molecular properties directly from the raw molecular geometry. This dexterous approach results from a combination of multiple factors such as the choice of ML model and an appropriate representation of the physical properties and invariance structure of molecules. In this work, a number of well established deep learning methods have been evaluated for predicting the molecular properties. In addition, the influence of these methods on various molecular representations are examined. The results of this evaluation show that deep learning methods can be successfully applied for accelerating the chemical accuracy of molecular properties. The application of deep learning methods on quantum-chemical systems especially for the purpose of molecular properties prediction, is a general direction that is well worth for pursuing a future research.