20.07.2013 Views

Raport de cercetare - Lorentz JÄNTSCHI

Raport de cercetare - Lorentz JÄNTSCHI

Raport de cercetare - Lorentz JÄNTSCHI

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

• Popularity<br />

• Universal Approximators<br />

• Nonlinearity<br />

• Robustness<br />

• Explanatory Knowledge<br />

Software<br />

R statistical environment<br />

Supervised Learning<br />

Handling Missing Data<br />

Outliers<br />

Non numerical variable remapping<br />

Feature Selection<br />

Classification Metrics<br />

Receiver Operating Characteristic (ROC)<br />

Regression Metrics<br />

Validation method: how to estimate the performance?<br />

MLP Training Algorithm<br />

Local Minima with MLP<br />

Overfitting<br />

Case Study: Intensive Care Medicine (Classification)<br />

Knowledge extraction (Decision Tree example for the renal organ)<br />

Case Study: Lamb Meat Quality (Regression)<br />

o Functional Networks<br />

Introduction to functional networks<br />

Differences between functional and artificial neural networks<br />

Functional equations<br />

Working with functional networks<br />

Different mo<strong>de</strong>ls<br />

Applications<br />

o Multi-class SVMs, Theory<br />

Yann Guermeur<br />

Guaranteed risk for large margin multi-category classifiers<br />

• Theoretical framework<br />

• Basic uniform convergence result<br />

• γ-ψ-dimensions<br />

• Generalized Sauer-Shelah lemma<br />

• Nature and rate of convergence<br />

Multi-class SVM<br />

• Multi-cathegory classification with binary SVMs<br />

• Class of functions implemented by the M-SVMs<br />

• Generalized formulation of the training algorithm<br />

• Three main mo<strong>de</strong>ls of M-SVMs<br />

• Some variants of the main mo<strong>de</strong>ls<br />

• Margins and support vectors<br />

Guaranteed risk for multi-class SVMs<br />

• Bounds on the covering numbers<br />

• Use of the Ra<strong>de</strong>macher complexity<br />

Mo<strong>de</strong>l selection for multi-class SVMs<br />

• Algorithms fitting the entire regularization path<br />

• Bounds on the leave-one-aut cross-validation error<br />

Open problems<br />

138

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!