Speaker
Description
Mathematical models of biology commonly use differential equation formulations. Certain application areas, such as signal transduction modeling or scientific machine learning, involve models that contain many parameters. Efficient training of these models requires sensitivity analysis that scales well as the number of parameters grows. Hence, adjoint sensitivity analysis (ASA) is typically employed, instead of forward sensitivity analysis (FSA).
In this work, we derive a new sensitivity analysis method that has similar scaling properties to ASA but, unlike ASA, can be solved in the forward direction. This provides some computational efficiency gains in terms of memory and complexity, especially for the stiff systems that are common when modeling biology. Furthermore, higher-order sensitivities are cheaper to compute with the new method. A drawback is that, when the parameter size is small or the state size is large, then the FSA or ASA methods, respectively, can naïvely outperform the new method.
Bibliography
@article{Fröhlich_Kaltenbacher_Theis_Hasenauer_2017, title={Scalable Parameter Estimation for Genome-Scale Biochemical Reaction Networks}, volume={13}, ISSN={1553-7358}, url={https://dx.plos.org/10.1371/journal.pcbi.1005331}, DOI={10.1371/journal.pcbi.1005331}, number={1}, journal={PLOS Computational Biology}, author={Fröhlich, Fabian and Kaltenbacher, Barbara and Theis, Fabian J. and Hasenauer, Jan}, editor={Stelling, Jorg}, year={2017}, month=jan, pages={e1005331}, language={en} }
