Robust Logistic Regression

• misclassified outliers can greatly affect models
• want: upper bound on regression coefficients
• first idea: linear combination of uninformative prior and logistic regression

$$p(y|x) = \pi \text{Ber}(y|0.5) + (1-\pi) \text{Ber}(y | \sigma(w^{T}x + b))$$