A New Adaptive Feedback Learning Algorithm for Multitask Learning
Zhenxing Li, Yong Qi, Weihua Li · 2012
Within a learning machine, we could improve the accuracy of every learning task while a few tasks learned together. This method is called multitask learning. It is popular to using simulated annealing algorithm or setting a constant as learning rate in multitask learning. Either is simple and efficiently. But neither use feedback information in machine effectively. A novel adaptive feedback learning algorithm (AFLA) for multitask learning is proposed in this paper. The AFLA is proposed to address the problem using feedback information in multitask learning machine. The AFLA uses the error signal as feedback information. And it applies error signal in learning rate. The experiments are carried on two datasets. With simulated annealing and constant comparisons, experimental results show that the AFLA can use error signal efficiently. At normal learning rate, the accuracy of AFLA is better than using simulated annealing algorithm and constant. At smaller learning rate, the accuracy of AFLA is 2-5 times as big as others.