Statistical Learning Theory and Applications
A graduate course on supervised learning theory, covering kernels, support vector machines and generalization bounds, with applications in vision and biology.
When can a function learned from a few examples be trusted on new data? This graduate course answers through statistical learning theory, starting with approximating functions in many variables from sparse data. It builds regularization and support vector machines, derives generalization bounds, and then shows the ideas at work in vision, speech and biology through guest lectures and projects.
What You Will Achieve
- Use regularization for regression and classification problems
- Derive generalization bounds with stability and VC theory
- Explain boosting, ranking and active learning at a research level
- Assess how learning methods apply to vision and bioinformatics
Program Overview
Kernels and Support Vector Machines
After an overview of the learning problem, the course develops its central tools.
- Reproducing kernel Hilbert spaces
- Regression and least-squares classification
- Support vector machines
- Manifold regularization
Generalization Bounds and Stability
A theoretical stretch asks how well a learned function will do on unseen examples.
- Stability of Tikhonov regularization
- Uniform convergence
- VC-dimension
- Rademacher averages
Applications and Project Presentations
Guest lecturers show the methods in use, and students present their own final projects.
- Computer vision object detection
- Speech and audio
- Morphable models for video
- Bioinformatics
Course Completion
Job Outlook
This background supports research in machine learning, computational neuroscience and computer vision, along with scientific work that builds predictive models.
Good for: Graduate students in computational neuroscience or machine learning who want the theory of kernels, regularization and generalization, taught with applications in view.
Keep in mind: The class dates from 2006, so its research topics reflect that year. Only two problem sets are published, and there are no solutions, so practice is limited.
Bottom line: A useful record of kernel-era learning theory, stronger as lecture notes to study than as a self-contained workout.
Frequently Asked Questions
What are the math camp sessions?
Two extra sessions review functional analysis and probability theory. They sit outside the numbered schedule and support the mathematics used throughout the course.
Which projects were students offered?
A published list suggests topics such as feature selection for support vector machines, kernel synthesis, local versus global classifiers and a history of induction. Students could also propose their own.
Who teaches the application sessions?
Guest lecturers cover ranking, object detection, online learning, neuroscience, speech, active learning, morphable models for video and bioinformatics, alongside the regular lecturers who teach the theory.



There are no reviews yet.