压缩回归学习算法的泛化界

Generalization Bounds of Compressed Regression Learning Algorithm

研究了压缩最小平方回归学习算法的泛化性问题.利用随机投影、覆盖数等理论以及概率不等式得到了该学习算法的泛化误差上界.所获结果表明:压缩学习虽以增大逼近误差的方式降低样本误差,但其增量是可控的.此外,通过压缩学习,在一定程度上克服了学习过程中所出现的过拟合现象.

This paper addresses the generalization performance of compressed least-square regression learning algorithm. A generalization error bound of this algorithm is established by using the random projection and the theory of the covering number. The obtained results show that the compressed learning can reduce the sample error a.t the price of increasing the approximation error, but the increment can be controlled. In addition, by using compressed projection, the overfitting problem for the learning algorithm can be overcome to a certain extent.