Landscape-Aware Growing: The Power of a Little LAG

TL;DR

This paper introduces a landscape-aware growing (LAG) approach that uses early training dynamics to better select optimal model growing strategies, improving efficiency in training Transformer models.

Contribution

It proposes a novel perspective based on early training behavior, moving beyond initialization metrics, and develops an adaptive strategy for model stacking.

Findings

Early training performance correlates better with final results than initialization metrics.

LAG enables more accurate prediction of optimal growing strategies.

Adaptive stacking strategies improve training efficiency.

Abstract

Recently, there has been increasing interest in efficient pretraining paradigms for training Transformer-based models. Several recent approaches use smaller models to initialize larger models in order to save computation (e.g., stacking and fusion). In this work, we study the fundamental question of how to select the best growing strategy from a given pool of growing strategies. Prior works have extensively focused on loss- and/or function-preserving behavior at initialization or simply performance at the end of training. Instead, we identify that behavior at initialization can be misleading as a predictor of final performance and present an alternative perspective based on early training dynamics, which we call "landscape-aware growing (LAG)". We perform extensive analysis of correlation of the final performance with performance in the initial steps of training and find early and more accurate predictions of the optimal growing strategy (i.e., with only a small "lag" after initialization). This perspective also motivates an adaptive strategy for gradual stacking.