Optimistic Adaptive Accelaration For Optimization on CIFAR-10 dataset🐶🐈🚘✈️ for image classification - Computer Vision
Predicting gradients beforehand will greatly reduce the number of epochs to be trained. Combining Optimistic Online Learning with adaptivity and the momentum to create the OPTIMISTIC-AMSGrad is a good idea. The implementation of the algorithm shows that OPTIMISTIC-AMSGRAD improves AMSGRAD in terms of various measures: training loss, testing loss, and classification accuracy on training/testing data over epochs. The basis of this algorithm is optimistic online learning. The basic idea behind online learning is to have a good guess over the loss function before choosing action and then the learner should exploit the guess to choose an action.