In this thesis, we address the problem of uncertainty quantification of supervised learning using the conditional generative model.
Our goal is to construct a conditional generative model that can capture predictive uncertainty while being theoretical...
In this thesis, we address the problem of uncertainty quantification of supervised learning using the conditional generative model.
Our goal is to construct a conditional generative model that can capture predictive uncertainty while being theoretically well-grounded.
To this end, we use the conditional generative adversarial network(cGAN) and develop two extensions of cGAN for uncertainty quantification of supervised learning: (1) a $k$ nearest neighborhood($k$NN) based data augmentation that enriches local variability around each sample, and (2) an uncertainty regularization term that explicitly enforces stochastic diversity in the generator outputs.
On the theoretical side, we establish asymptotic guarantees for the conditional density estimated by cGAN.
Specifically, we show that the conditional density estimated by a conditional generative adversarial network converges to the true conditional density in Jensen-Shannon divergence at a nearly minimax-optimal rate.
We further show that the bias introduced by the $k$NN-based data augmentation converges to zero faster than the original convergence rate when $k$ increases at an appropriate rate, and the proposed uncertainty regularization preserves the optimal convergence rate when the regularization coefficient vanishes at an appropriate rate.
Through experiments on both regression and classification datasets, we empirically validate the proposed framework.
The results demonstrate that our methods achieve competitive or superior performance compared with established uncertainty quantification baselines.
Taken together, these findings indicate that conditional generative adversarial networks, when properly regularized or augmented, constitute a powerful and theoretically grounded tool for uncertainty quantification in supervised learning.