GSoC '26

I will post any updates here on this page.

I might add more details later in order to make these more useful than just a simple log of activities.

Week 1-2

So far I have a good handle on how the package MathOptAI.jl works and how the workflow of implementing a new predictor should be handled. I have implemented ReLUEpigraph and currently testing it on a practical problem.

I have also developed a Flux model which is an Input Convex Neural Network (ICNN) that can be embedded in a JuMP.Model and using ReLUEpigraph enables the ICNN to be represented with an LP formulation. The PR for the ReLUEpigraph is created here.

Week 3-4

Currently working on conic reformulations and more epigraph formulations. Activation functions like softplus can be represented with exponential cones and then we can use solvers like MOSEK or SCS to solve the conic problems.

I have also submitted a PR for the Flux tutorial of ICNN. It can be viewed here.

Week 5-6

This week, we have added another ICNN tuorial, this time using PyTorch. The PR for it can be viewed here. We have also added a small training loop to both Flux tutorial and PyTorch tutorial to show the fit on a convex function.

We have also finalized and added SoftPlusConicEpigraph with test using SCS. The PR for it can be viewed here.