Jia Hao (Frank) Liang

@MIT '29, 6-4&18
@MIT_CSAIL UROP

[Jia Hao (Frank) Liang @flbjcy]

Very glad to have worked with Xingjian Bai (@SimulatedAnneal), Professor Vincent Sitzmann (@vincesitzmann), and Jayden Teoh (@jayden_teoh_). Excited for what comes next, and always happy to chat about the work!

[Jia Hao (Frank) Liang @flbjcy]

Optimizers—one of the most fundamental backbones in ML 101—can now be reactive?!

In the past, from fixed schedules to adaptive optimizers, we were using a fixed set of operations with predefined if/else conditions to solve a fundamentally dynamic problem.

Today, we present OPAL

[Xingjian Bai @SimulatedAnneal]

1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?

We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.

In every language and image setting we tried that has no tuned recipe, it matches or beats the best recipe one can one-shot, and it guards against loss spikes and overfitting by itself. In the near future, optimization decisions may belong to agents, not humans.

📝 https://t.co/augwW3iFsB
💻 Try it on your task: https://t.co/NdQNekaHoc
🧵

[Jia Hao (Frank) Liang @flbjcy]

What makes OPAL reactive? In a nutshell, it can learn from training signals and improve its own optimizer code, opening the door to effective optimizers no one has tried before! Our experiments show that OPAL performs well across different settings.