Calibrated Dataset Condensation for Faster Hyperparameter Search
Mucong Ding, Yuancheng Xu, Tahseen Rabbani, Xiaoyu Liu, Brian Gravelle, Teresa Ranadive, Tai-Ching Tuan, Furong Huang
TL;DR
The paper tackles the high cost of hyperparameter search by proposing Hyperparameter Calibrated Dataset Condensation (HCDC), which preserves hyperparameter optimization outcomes rather than single-model generalization. By aligning hypergradients computed on condensed data with those from the full dataset, and using implicit differentiation with a Neumann-series inverse-Hessian, HCDC learns a synthetic validation set that mirrors hyperparameter rankings across architectures and hyperparameters. The method enables scalable, memory-efficient hypergradient computation and extends the hyperparameter space to handle discrete and continuous settings, achieving strong empirical preservation of NAS rankings on both image and graph tasks, while accelerating search. This approach provides a practical pathway to faster, data-efficient hyperparameter and architecture search with broad applicability to vision and graph learning problems.
Abstract
Dataset condensation can be used to reduce the computational cost of training multiple models on a large dataset by condensing the training dataset into a small synthetic set. State-of-the-art approaches rely on matching the model gradients between the real and synthetic data. However, there is no theoretical guarantee of the generalizability of the condensed data: data condensation often generalizes poorly across hyperparameters/architectures in practice. This paper considers a different condensation objective specifically geared toward hyperparameter search. We aim to generate a synthetic validation dataset so that the validation-performance rankings of the models, with different hyperparameters, on the condensed and original datasets are comparable. We propose a novel hyperparameter-calibrated dataset condensation (HCDC) algorithm, which obtains the synthetic validation dataset by matching the hyperparameter gradients computed via implicit differentiation and efficient inverse Hessian approximation. Experiments demonstrate that the proposed framework effectively maintains the validation-performance rankings of models and speeds up hyperparameter/architecture search for tasks on both images and graphs.