[2L4-GS-1] Fundamental AI, theory, algorithm:

[2L4-GS-1-05] Learning Nonlinear Regression Model from Two-Sample Averaged Data

〇Ryuta Matsuno¹, Akira Kitaoka¹, Keita Sakuma¹, Masakazu Hirokawa¹ (1. NEC Corporation)

Keywords:Learning from Averaged Data, Learning from Aggregated Data, Privacy-Preserving Machine Learning

Data averaged over multiple samples is promising for privacy-preserving machine learning, as it retains much information for machine learning while protecting privacy. As a first step in this field, this study proposes a method for learning a general nonlinear regression model from two-sample averaged data. The proposed method trains a nonlinear regression model by maximizing our efficiently approximated likelihood. Experiments on synthetic data demonstrate that the prediction performance of our method is even competitive with that of the oracle when the input dimension is small. Moreover, experiments on real-world data confirm that the proposed method effectively trains neural networks, achieving superior results compared to the baseline methods. This research opens up a new direction for privacy-preserving machine learning.