Abstract
Deep unrolling is a promising technique for the practical deployment of deep neural networks hindered by their black-box nature. The parameters of the network can be solved by various optimizers such as the inertial proximal alternating linearized minimization. Since the loss function of unrolled network presents a highly non-convex shape with many local minima, unlike the convex learning invariant to the training order, the order of training data affects the deep-unrolled learning performance and may degrade the final solution. To solve this problem, a data scheduling strategy under the curriculum learning framework is proposed. Curriculum learning imitates the meaningful learning order in human curricula from simple concepts to hard problems. Based on a novel scoring function on image groups, training data are ranked with the difficulty measurer. The curriculum and anti-curriculum schedulers are created using the learning priority. The proposed data scheduler is applied on magnetic resonance image reconstruction which is compared to the reconstruction results without using any curriculum learning-based training strategy. Experimental results show that the proposed method is superior to the conventional method in terms of the root mean squared error and the structural similarity index. The pacing function will be investigated in future work.