Session 10- PAC Theory and Generalization Bound | Topics in AI/ML-Concepts & Applications скачать в хорошем качестве

Session 10- PAC Theory and Generalization Bound | Topics in AI/ML-Concepts & Applications

In this session we study the PAC Theory (Probably Approximately Correct Framework) in detail. This is an important concept under the overall Theory of Learning, and we explain it in a manner that makes it easy to understand for an undergraduate engineer. We develop the theory starting with the basic Hoeffding Inequality leading to the Vapnik Chervonenkis Bound and then make our way to the Generalization bound. The concepts of Growth Functions and VC Dimension (covered in an earlier lecture) are utilized here. Every idea which goes into the PAC Theory is described in detail, with analogies and examples which help the viewer relate to every day life. We also discuss the definition of PAC Learnability which results from this theory. We finally offer the practical implications of the PAC Theory including the relationship between the Number of training samples and the VC Bound. 5:10 - Correlating Eout and Ein 12:11 - PAC Framework 21:35 - Hoeffding's Inequality 33:03 - Extension to Hypothesis Sets 54:21 - VC Bound (Leverages results from previous lecture) 1:02:40 - Generalization and Learning 1:11:32 - Practical Implications 1:24:24 - PAC Learnable Problem 1:28:35 - Generalization Bound 1:32:30 - Summary