论文部分内容阅读
多分类学习中经常需要考虑在泛化性能和计算开销间进行权衡。本文提出一个生成式概率多分类器,综合考虑了泛化性和学习/预测速率。我们首先证明了我们的分类器具有最大间隔性质,这意味着对于未来数据的预测精度几乎和训练阶段一样高。此外,我们消除了目标函数中的大量的局部变元,极大地简化了优化问题。通过凸分析和概率语义分析,我们设计了高效的在线算法,与经典情形的最大不同在于这个算法使用聚集而非平均化处理梯度。实验证明了我们的算法具有很好的泛化性能和收敛速度。
It is often necessary to consider trade-offs between generalization and computational overhead in multi-category learning. This paper presents a generative probabilistic multi-classifier that takes into account generalization and learning / prediction rates. We first prove that our classifier has the largest interval property, which means that the prediction accuracy for future data is almost as high as the training phase. In addition, we eliminate a large number of local variables in the objective function, which greatly simplifies the optimization problem. By convex analysis and probabilistic semantic analysis, we designed an efficient online algorithm that is most different from the classic case where the algorithm uses aggregation instead of averaging the gradients. Experiments show that our algorithm has good generalization performance and convergence speed.