UC San Diego

Machine learning has consistently proved to be useful in many applications. An integral facet allowing for machine learning is the training step. The training step of the models may take up excessive time and hardware resources. This is especially relevant when considering the fact that the performance of the models depend heavily on the aptitude of the training step. Depending on the scale of the application, any increase in performance during the training step may pay large dividends regarding the efficiency and success of the model as a whole. A heuristic commonly implemented as part of the training steps within models is the ordering of training samples with some predefined knowledge, more popularly known as curriculum learning. The common assumption resides in the belief that applying curriculum learning will generally improve results. In this study, experiments were performed through the utilization of curriculum learning, while isolating certain variables to analyze the effect of curriculum learning as a whole. Within the experiments, different metrics of hardness will be explored. The training process will train on samples which will be ordered based on the different metrics of hardness. The orderings typically covered are Easy-to-Hard, Hard-to-Easy, and Random. Easy-to-Hard matches the analogy of how humans and animals typically learn. In Easy-to-Hard, the learner is presented with the easiest samples followed by progressively harder samples. In Hard-to-Easy, the learner is presented with the hardest samples followed by progressively easier samples. In Random, the learner is presented with samples that maintain no inherent ordering. Results obtained during this study reveal that curriculum learning does indeed have an effect on performance, but the effects can both be positive or negative, depending on the the metric of hardness. Such results garnered evidence that, when selecting a curriculum strategy for a setting, one should use care as the effectiveness of the training can be either diminished or enhanced based on the metric of hardness chosen.