PhD student at École des Ponts (France), working on machine learning and operations research with applications to railway planning.
We present a Julia package for differentiating through functions that are defined implicitly. It can be used to compute derivatives for a wide array of "black box" procedures, from optimization algorithms to fixed point iterations or systems of nonlinear equations.
Since it mostly relies on defining custom chain rules, our code is lightweight and integrates nicely with Julia's automatic differentiation and machine learning ecosystem.
We present InferOpt.jl, a generic package for combining combinatorial optimization algorithms with machine learning models. It has two purposes:
- Increasing the expressivity of learning models thanks to new types of structured layers.
- Increasing the efficiency of optimization algorithms thanks to an additional inference step.
Our library provides wrappers for several state-of-the-art methods in order to make them compatible with Julia's automatic differentiation ecosystem.