Montag, 13.12.2021, 11.00 Uhr

Global Derivatives



In this talk, we show how to obtain global derivatives that are guaranteed enclosures of the derivative information on specified domains. Therefore, we combine algorithmic differentiation (AD) methods with interval arithmetic and McCormick relaxations. While naive interval computations are prone to overestimation of exact value ranges, we identify special cases for which the natural interval extension applied to the AD methods compute exact value ranges for the global derivatives.

We present two applications that benefit from global derivatives: deterministic global optimization by branch-and-bound methods, and significance-driven unreliable and approximate computing. Within the framework of the global optimization case study we introduce subdomain separability. This local property enables the partitioning of the optimization problem on subdomains that fulfill a certain monotonicity condition. The approximate computing case study demonstrates how to automatically prune artificial neural networks by using significance values.


Es laden ein: die Dozentinnen und Dozenten der Informatik