ECE 299: Vapnik-Chervonenkis classes

Having covered generalization bounds for abstract ERM via Rademacher averages, we now trace the historical development of the field:

  • Vapnik-Chervonenkis classes: shatter coefficients; VC dimension; examples of VC classes; Sauer-Shelah lemma; implication for Rademacher averages

Mmmmm, character-building!

Statistical Learning Theory (ECE 299, Spring 2011)

Now that the ITA Workshop is over and the ISIT deadline is officially past, I can resume semi-regular blogging.

This semester, I am teaching a graduate course on Statistical Learning Theory in my department. My aim is to introduce graduate students in electrical engineering to such things as Empirical Risk Minimization, generalization bounds, model selection, complexity regularization, minimax lower bounds &c., with certain examples of applications to information theory, signal processing, and adaptive control.

Unfortunately, I have not come across a textbook that would be suitable for teaching statistical learning theory to graduate students of engineering. Most texts out there are skewed towards either the computer science side of things or the mathematical statistics crowd. The only book that comes somewhat close to what I had in mind is M. Vidyasagar‘s Learning and Generalization, but its $169 price tag has stopped me from officially adopting it. Instead, I have been preparing my own lecture notes. As the class goes on, I will be posting additional notes here, but here is what I have covered so far:

  • Introduction: learning from examples in a probabilistic setting; goals of learning; basics of statistical decision theory; estimation and approximation errors
  • Concentration inequalities: Markov’s and Chebyshev’s inequalities; the Chernoff bounding trick; Hoeffding’s inequality; McDiarmid’s inequality; examples
  • Formulation of the learning problem: concept and function learning in the realizable case; PAC learning; model-free learning; Empirical Risk Minimization and its consistency
  • More on Empirical Risk Minimization: error bounds via Rademacher averages

Normally, I write these in frenetic two-hour bursts, so they are more than likely full of misconceptions, omissions, appalling lack of rigor, bugs, typos, and the like. Caveat lector!