Probably Approximately Correct
1984年,Leslie Valiant在A Theory of the Learnable提出的一种理论框架,结合计算复杂度理论中的方法对学习算法的可行性进行分析。
Definition - Realizable Case
A concept class is said to be PAC-learnable if there exists an algorithm and a polynomial function such that for any and , for all distributions on and for any target concept , the following holds for any sample size :
If further runs in , then is said to be efficiently PAC-learnable. When such an algorithm exists, it is called a PAC-learning algorithm for .
特点
- PAC Learning中学习算法被动地接受样本(一种passive learning),样本从某个任意但未知的分布中获得。PAC可学习性要求算法应该在所有分布下满足要求。
- 泛化误差被定义为假设与概念间结论不一致的概率,这个概率相对于同样的分布。PAC Learning的训练和测试都是在同一个分布下
- 是泛化误差的界,则衡量信心。由于分布是任意的,总是存在极端情况,因此PAC只要求有一定的概率近似正确,这一概率随着样本量的增加而增加。
- n代表表示一个样本需要的cost,比如欧几里得空间的维度
Definition - Agnostic Case/Agnostic PAC Learning
Let be a hypothesis set. is an agnostic PAC-learning algorithm if there exists a polynomial function such that for any and , for all distributions over , the following holds for any sample size :
If further runs in , then it is said to be an efficient agnostic PAC-learning algorithm.
对比 & 联系
- agnostic learning 蕴涵 realizable learning,反之亦然1。
- agnostic learning的目标是学习到的假设尽量接近误差最小的假设
- agnostic learning通常不是deterministic的,而是stochastic的,故用上的联合分布来表示。如果每个样本的标签是确定的,因为标签函数对应的假设的误差是0,只要它在假设类里,那么可以转变为realizable的情况。