Online Bayesian Passive-Aggressive Learning

  title={Online Bayesian Passive-Aggressive Learning},
  author={Tianlin Shi and Jun Zhu},
  booktitle={Journal of Machine Learning Research},
where the constant C = exp(ytτtμt xt + 12τ t xt xt) . Therefore, the distribution qt+1(w) = N (μt + τtxt, I) and as a by-product, the normalization term Γ(τt) = √ 2π K exp(τtytx > t μt + 1 2τ 2 t x > t xt). Next, we show that μt+1 = μt + τytxt is the optimal solution of the online Passive-Aggressive update rule (Crammer et al., 2006). To see this, we plug the derived Γ(τt) into (26), ignore constant terms and obtain 

From This Paper

Topics from this paper.

Similar Papers

Loading similar papers…