Tri-self-taught learning of artificial neural networks

We present a conceptually simple and general framework for self-taught learning, and this method can modify weights of neural networks when making prediction on samples according to the previous knowledge. The method called Tri-STNN. Based on Tri-training, it adds a Judge Network to make a higher accuracy of prediction. Tri-STNN is simple to be trained and the training datasets for Judge Network is easy to obtain. Moreover, Tri-STNN is capable of lots of tasks. We test Tri-STNN on the datasets of CIFAR-10, and the result shows that Tri-STNN can constantly keep self-taught learning and improve the generalization ability for new samples.