| Author | Anna Novokshonova |
| Consultants | Fedor Sobolevsky Muhammadsharif Nabiev |
| Advisor | Oleg Bakhteev, PhD |
This paper investigates inductive bias in machine learning models. By inductive bias we mean the preference of a model for certain types of functions or data structures over others. To analyze the inductive bias of a fixed model, we consider a problem of finding data that this model can fit and generalize on particularly well. Previous work demonstrated that generating labels for a fixed dataset allows one to extract the inductive bias. Here, we extend this approach by proposing a method for generating full synthetic datasets. We train a generative model to produce datasets on which the target model achieves strong generalization performance. We test the proposed framework on CNN and RNN, and analyze obtained datasets for each model.
If you find our work helpful, please cite us.
@article{citekey,
title={Title},
author={Name Surname, Name Surname (consultant), Name Surname (advisor)},
year={2025}
}Our project is MIT licensed. See LICENSE for details.