Curriculum Learning for Depth Estimation with Deep Convolutional Neural Networks
Ajay Surendranath, Dinesh Babu Jayagopi · 2018
Curriculum learning is a machine learning technique adapted from the way humans acquire knowledge and skills, initially mastering simple tasks and progressing to more complex tasks. The work explores curriculum training by creating multiple levels of dataset with increasing complexity on which the trainings are performed. The experiments demonstrated that there is an average of 12% improvement test loss when compared to a non-curriculum approach. The experiment also demonstrates the advantage of creating synthetic dataset and how it aids in the overall improvement of accuracy. An improvement of 26% is attained on the test error loss when curriculum trained model was compared to training on a limited real world dataset. The work also goes onto propose a novel learning approach, the Self Paced Learning approach with Error-Diversity (SPL-ED) An overall reduction of 32% in the test loss is observed when compared to the non-curriculum training limited to real-world dataset.