Practical tradeoffs between memory, compute, and performance in learned optimizers

TL;DR

This paper analyzes the trade-offs between memory, compute, and performance in learned optimizers, and introduces a new optimizer that is faster and more memory-efficient.

Contribution

It identifies key design features affecting resource trade-offs and develops a learned optimizer that improves on speed and memory efficiency.

Findings

Quantifies memory, compute, and performance trade-offs in optimizers.

Develops a learned optimizer that is faster and more memory-efficient.

Provides open-source code for the proposed optimizer.

Abstract

Optimization plays a costly and crucial role in developing machine learning systems. In learned optimizers, the few hyperparameters of commonly used hand-designed optimizers, e.g. Adam or SGD, are replaced with flexible parametric functions. The parameters of these functions are then optimized so that the resulting learned optimizer minimizes a target loss on a chosen class of models. Learned optimizers can both reduce the number of required training steps and improve the final test loss. However, they can be expensive to train, and once trained can be expensive to use due to computational and memory overhead for the optimizer itself. In this work, we identify and quantify the design features governing the memory, compute, and performance trade-offs for many learned and hand-designed optimizers. We further leverage our analysis to construct a learned optimizer that is both faster and more memory efficient than previous work. Our model and training code are open source.