A DPI-PAC-Bayesian Framework for Generalization Bounds
By: Muhan Guan, Farhad Farokhi, Jingge Zhu
Potential Business Impact:
Makes computer learning more accurate and reliable.
We develop a unified Data Processing Inequality PAC-Bayesian framework -- abbreviated DPI-PAC-Bayesian -- for deriving generalization error bounds in the supervised learning setting. By embedding the Data Processing Inequality (DPI) into the change-of-measure technique, we obtain explicit bounds on the binary Kullback-Leibler generalization gap for both R\'enyi divergence and any $f$-divergence measured between a data-independent prior distribution and an algorithm-dependent posterior distribution. We present three bounds derived under our framework using R\'enyi, Hellinger \(p\) and Chi-Squared divergences. Additionally, our framework also demonstrates a close connection with other well-known bounds. When the prior distribution is chosen to be uniform, our bounds recover the classical Occam's Razor bound and, crucially, eliminate the extraneous \(\log(2\sqrt{n})/n\) slack present in the PAC-Bayes bound, thereby achieving tighter results. The framework thus bridges data-processing and PAC-Bayesian perspectives, providing a flexible, information-theoretic tool to construct generalization guarantees.
Similar Papers
A DPI-PAC-Bayesian Framework for Generalization Bounds
Information Theory
Makes computer learning more accurate and reliable.
Generalization Bounds for Quantum Learning via Rényi Divergences
Quantum Physics
Makes smart computers learn better from less data.
How good is PAC-Bayes at explaining generalisation?
Machine Learning (Stat)
Helps computers learn better with fewer mistakes.