A Survey on Curriculum Learning
Abstract : ? Curriculum learning (CL) is a training strategy that trains a machine learning model from easier data to harder data, which imitates the meaningful learning order in human curricula.
? As an easy-to-use plug-in, the CL strategy has demonstrated its power in improving the generalization capacity and convergence rate of various models in a wide range of scenarios such as computer vision and natural language processing etc.
? In this survey article, we comprehensively review CL from various aspects including motivations, definitions, theories, and applications.
? We discuss works on curriculum learning within a general CL framework, elaborating on how to design a manually predefined curriculum or an automatic curriculum.
?
? In particular, we summarize existing CL designs based on the general framework of Difficulty Measurer + Training Scheduler and further categorize the methodologies for automatic CL into four groups, i.e., Self-paced Learning, Transfer Teacher, RL Teacher, and Other Automatic CL.
? We also analyze principles to select different CL designs that may benefit practical applications.
? Finally, we present our insights on the relationships connecting CL and other machine learning concepts including transfer learning, meta-learning, continual learning and active learning, etc., then point out challenges in CL as well as potential future research directions deserving further investigations.
? Our first contribution is to formalize the existing curriculum learning methods under a single umbrella.
? This enables us to define a generic formulation of curriculum learning.
? Additionally, we link curriculum learning with the four main components of any machine learning approach: the data, the model, the task and the performance measure.
? We observe that curriculum learning can be applied on each of these components, all these forms of curriculum having a joint interpretation linked to loss function smoothing.
? we can categorize the motivations for applying CL into two groups: to guide, regularizing the training towards better regions in parameter space (with steeper gradients) as from the perspective of the optimization problem, and to denoise, focusing on high-confidence easier area to alleviate the interference of noisy data as from the perspective of data distribution.
? Fortunately, the AOS algorithm naturally decomposes SPL into two problems of optimizing w and v, which makes it feasible to embed the loss prior knowledge into SPL by encoding it as a part of SP-regularizer or a constraint on v
? The proposed categories stem from the different assumptions, model requirements and training methodologies applied in each work.
? A self-paced dictionary learning method for image classification is proposed by Tang et al.
? They employ an easy-to-hard approach which introduces information about the complexity of the samples into the learning procedure.
? The easy examples are automatically chosen at each iteration, using the learned dictionary from the previous iteration, with more difficult samples being gradually introduced at each step.
? The advantages of applying CL training strategies to miscellaneous real-world scenarios can be mainly summarized as improving the model performance on target tasks and accelerating the training process, which cover the two most significant requirements in major machine learning research.
? The most valuable advantages of SPL over predefined CL are mainly two-fold: 1) SPL is semi-automatic CL with a loss-based automatic Difficulty Measurer and dynamic curriculum, which makes it more flexible and adaptive for various tasks and data distributions. 2) SPL embeds the curriculum design into the learning objective of the original machine learning tasks, which makes it widely applicable as a plug-in tool.
|