Hill-ADAM Optimizer Machine Learning

[2manikan]

Select Topic Area

Show & Tell

Body

Hi everyone, I'm so excited to share my latest research project over the summer, in which I created a novel Machine Learning optimizer (Hill-ADAM) for prescribed loss functions! The optimizer explores the loss space in search of the global minimum by alternating between minimization and maximization (see gif). I have learned a lot and was able to hone my skills through this research project thanks to my amazing advisor Dr. Leilani Gilpin!

Check it out here:
Arxiv: https://www.arxiv.org/abs/2510.03613
Github: https://github.com/2manikan/Hill-ADAM

While the optimizer has a wide variety of possible applications, we applied the optimizer to perform source-to-target image color correction (with additional constraints to ensure realistic images). I am proud to say that we were able to achieve this goal!

[2manikan]

Hi, sorry about the late reply, I just saw this! Thanks so much for taking the time to check out Hill-ADAM! I'd love to check out your projects as well, they seem super interesting :) Best, Meena.

…

On Thu, Jan 8, 2026 at 12:35 PM JUAN CARLOS ***@***.***> wrote: This is really impressive work, congrats on Hill-ADAM! The idea of alternating between minimization and maximization to better explore the loss landscape is super interesting, especially seeing it applied to image color correction with realistic constraints. I’ll take some time to read the arXiv paper and go through the repo in more detail. If you ever feel like checking out other projects from the community, you’re very welcome to pass by my GitHub profile as well: https://github.com/jcduro — any feedback or ideas for improvement are greatly appreciated. 🚀 — Reply to this email directly, view it on GitHub <#183041 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BLZFDDA3TNMMAMGX6EZZLGL4F25PHAVCNFSM6AAAAACQGFRUISVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKNBUHA3TAMI> . You are receiving this because you authored the thread.Message ID: ***@***.***>