Model Performance Scaling with Multiple Data Sources

Tatsunori Hashimoto
Proceedings of the 38th International Conference on Machine Learning, PMLR 139:4107-4116, 2021.

Abstract

Real-world machine learning systems are often trained using a mix of data sources with varying cost and quality. Understanding how the size and composition of a training dataset affect model performance is critical for advancing our understanding of generalization, as well as designing more effective data collection policies. We show that there is a simple scaling law that predicts the loss incurred by a model even under varying dataset composition. Our work expands recent observations of scaling laws for log-linear generalization error in the i.i.d setting and uses this to cast model performance prediction as a learning problem. Using the theory of optimal experimental design, we derive a simple rational function approximation to generalization error that can be fitted using a few model training runs. Our approach can achieve highly accurate ($r^2\approx .9$) predictions of model performance under substantial extrapolation in two different standard supervised learning tasks and is accurate ($r^2 \approx .83$) on more challenging machine translation and question answering tasks where many baselines achieve worse-than-random performance.

Cite this Paper

BibTeX
@InProceedings{pmlr-v139-hashimoto21a, title = {Model Performance Scaling with Multiple Data Sources}, author = {Hashimoto, Tatsunori}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, pages = {4107--4116}, year = {2021}, editor = {Meila, Marina and Zhang, Tong}, volume = {139}, series = {Proceedings of Machine Learning Research}, month = {18--24 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v139/hashimoto21a/hashimoto21a.pdf}, url = {https://proceedings.mlr.press/v139/hashimoto21a.html}, abstract = {Real-world machine learning systems are often trained using a mix of data sources with varying cost and quality. Understanding how the size and composition of a training dataset affect model performance is critical for advancing our understanding of generalization, as well as designing more effective data collection policies. We show that there is a simple scaling law that predicts the loss incurred by a model even under varying dataset composition. Our work expands recent observations of scaling laws for log-linear generalization error in the i.i.d setting and uses this to cast model performance prediction as a learning problem. Using the theory of optimal experimental design, we derive a simple rational function approximation to generalization error that can be fitted using a few model training runs. Our approach can achieve highly accurate ($r^2\approx .9$) predictions of model performance under substantial extrapolation in two different standard supervised learning tasks and is accurate ($r^2 \approx .83$) on more challenging machine translation and question answering tasks where many baselines achieve worse-than-random performance.} }
Endnote
%0 Conference Paper %T Model Performance Scaling with Multiple Data Sources %A Tatsunori Hashimoto %B Proceedings of the 38th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2021 %E Marina Meila %E Tong Zhang %F pmlr-v139-hashimoto21a %I PMLR %P 4107--4116 %U https://proceedings.mlr.press/v139/hashimoto21a.html %V 139 %X Real-world machine learning systems are often trained using a mix of data sources with varying cost and quality. Understanding how the size and composition of a training dataset affect model performance is critical for advancing our understanding of generalization, as well as designing more effective data collection policies. We show that there is a simple scaling law that predicts the loss incurred by a model even under varying dataset composition. Our work expands recent observations of scaling laws for log-linear generalization error in the i.i.d setting and uses this to cast model performance prediction as a learning problem. Using the theory of optimal experimental design, we derive a simple rational function approximation to generalization error that can be fitted using a few model training runs. Our approach can achieve highly accurate ($r^2\approx .9$) predictions of model performance under substantial extrapolation in two different standard supervised learning tasks and is accurate ($r^2 \approx .83$) on more challenging machine translation and question answering tasks where many baselines achieve worse-than-random performance.
APA
Hashimoto, T.. (2021). Model Performance Scaling with Multiple Data Sources. Proceedings of the 38th International Conference on Machine Learning, in Proceedings of Machine Learning Research 139:4107-4116 Available from https://proceedings.mlr.press/v139/hashimoto21a.html.

Related Material