Learning with noise. Extension to regression

Learning theory with noise provides an interesting framework. Outliers are a real-world problem. A simple model of outliers leads to similar conclusions than with much the difficult malicious errors; moreover, it sounds more realistic than constant noise, CPCN noise and malicious errors. The bias introduced by margin methods using distances to avoid NP-completeness can be a real problem and that asymptotic empirical risk minimization could be important