Abstract

We propose to learn a curriculum or a syllabus for supervised learning withdeep neural networks. Specifically, we learn weights for each sample intraining by an attached neural network, called ScreenerNet, to the originalnetwork and jointly train them in an end-to-end fashion. We show the networksaugmented with our ScreenerNet achieve early convergence with better accuracythan the state-of-the-art rule-based curricular learning methods in extensiveexperiments using three popular vision datasets including MNIST, CIFAR10 andPascal VOC2012, and a Cartpole task using Deep Q-learning.

 

Quick Read (beta)