Automatic differentiation (AD) is an increasingly important part of the numerical computing technology stack. But while AD is simple enough in principle, implementations that are completely general, flexible and fast remain elusive, and instead we have a number of different systems that make difficult tradeoffs. This talk will explore this issue from the perspective of general program transformations, arguing that AD implementation is largely limited by current compilers, rather than these being issues with AD in itself. Modern compiler technology, as well as careful design of numerical computing languages themselves, can get us out of the rut.

Please use this link to attend the virtual seminar: https://bluejeans.com/989480913