In this talk I discuss Full Waveform Inversion as an example of large adjoint problems. This problem is relevant in seismic and medical imaging. I will discuss this in the context of Devito, a domain-specific language for automatically generating high-performance finite-difference solvers. Additionally, I will make deep dives into the implementation of automatic differentiation for this DSL, as well as the use of lossy compression to accelerate these computations.
Please use this link to attend the virtual seminar.