Search papers, labs, and topics across Lattice.
This paper critiques the reliability of scaling laws in small-scale experiments, revealing that hyperparameter sensitivity significantly affects model performance at lower scales. The authors demonstrate that while scaling laws do exist for small models, their benefits are often obscured due to the extensive hyperparameter tuning required to reach optimal performance. By developing a new methodology that emphasizes hyperparameter tuning, they successfully replicate large-scale findings regarding the placement of normalization layers in transformer architectures using small-scale data.
Hyperparameter tuning is the hidden key that unlocks the scaling laws in small models, revealing insights previously thought to be exclusive to larger architectures.
Scaling laws promised cost-effective experiments; six years later, they have yet to fully deliver. Instead, researchers have found them unreliable at small scales (starting at 4M parameters) and concluded that sizable models cannot be avoided. We show this is not the case: the confounding factor is hyperparameters. Small models are highly sensitive, but hyperparameter sensitivity fades with scale. This small-scale sensitivity makes scaling laws easy to miss because they only emerge on the fully tuned frontier, and reaching that frontier requires an extensive search far beyond what most ever run. By ablating the basic scaling law recipe, we show well-tuned hyperparameters matter more than any other ingredient. Further, we reveal why those hyperparameters become easier to find: as scale increases, the hyperparameter loss surface becomes lower dimensional. Nevertheless while scaling laws exist in small models, extrapolation hits statistical limitations. A holistic approach is required. Synthesizing our insights with the recent literature, we develop a new methodology for model-centric research and demonstrate it on a question that once took the field years to settle: where to place normalization layers in the transformer architecture. From small-scale experiments, we recover the large scale result: pre-normalization works better as models grow in size. With the right tools and a better understanding, small-scale experiments can deliver on scaling laws'long-awaited promise.
AI2