[Paper Review] Curriculum Learning: A Survey

Petru Soviany, Radu Tudor Ionescu+2 more|arXiv (Cornell University)|Jan 25, 2021

Machine Learning and Data Classification4 citations

TL;DR

This survey provides a comprehensive overview of curriculum learning (CL) in machine learning, formalizing its principles and linking it to core components of learning frameworks. It presents a multi-perspective taxonomy and hierarchical clustering of CL methods, identifies key challenges such as sample ranking and pacing, and proposes future directions including automated curriculum regulation and integration with alternative optimization algorithms like evolutionary search.

ABSTRACT

Training machine learning models in a meaningful order, from the easy samples to the hard ones, using curriculum learning can provide performance improvements over the standard training approach based on random data shuffling, without any additional computational costs. Curriculum learning strategies have been successfully employed in all areas of machine learning, in a wide range of tasks. However, the necessity of finding a way to rank the samples from easy to hard, as well as the right pacing function for introducing more difficult data can limit the usage of the curriculum approaches. In this survey, we show how these limits have been tackled in the literature, and we present different curriculum learning instantiations for various tasks in machine learning. We construct a multi-perspective taxonomy of curriculum learning approaches by hand, considering various classification criteria. We further build a hierarchical tree of curriculum learning methods using an agglomerative clustering algorithm, linking the discovered clusters with our taxonomy. At the end, we provide some interesting directions for future work.

Motivation & Objective

• To formalize and unify existing curriculum learning approaches under a single, generic framework.
• To address the limitations of curriculum learning, such as difficulty in ranking samples from easy to hard and selecting optimal pacing functions.
• To provide a structured taxonomy and hierarchical clustering of curriculum learning methods based on classification criteria.
• To identify underexplored research directions, particularly in emerging domains like vision transformers and medical imaging.
• To promote wider adoption of curriculum learning by clarifying its benefits and integration challenges in deep learning workflows.

Proposed method

• Develops a multi-perspective taxonomy of curriculum learning methods based on classification criteria such as data type, learning objective, and curriculum strategy.
• Applies agglomerative clustering to group similar curriculum learning methods and constructs a hierarchical tree linking clusters to the taxonomy.
• Proposes a generic formulation of curriculum learning that integrates it with the four main components of any machine learning approach: data, model, optimization, and evaluation.
• Reviews existing CL instantiations across diverse tasks, including object detection, image classification, and neural machine translation.
• Analyzes the trade-offs between curriculum learning and opposing strategies like hard example mining and anti-curriculum learning.
• Proposes future research directions, including automated curriculum regulation and coupling CL with non-SGD optimization methods such as evolutionary algorithms.

Experimental results

Research questions

• RQ1 How can curriculum learning be formally unified across diverse machine learning tasks and architectures?
• RQ2 What are the key challenges in implementing curriculum learning, particularly in sample ranking and pacing function design?
• RQ3 How does curriculum learning compare to alternative training strategies such as hard example mining or anti-curriculum learning?
• RQ4 In which emerging ML domains—such as vision transformers or medical imaging—does curriculum learning show the most promise?
• RQ5 What mechanisms can enable automated, adaptive curriculum scheduling that reduces manual intervention?

Key findings

• Curriculum learning consistently improves training convergence speed and final model accuracy across multiple domains, including computer vision, NLP, and speech recognition.
• The survey identifies that the main barrier to widespread adoption is the difficulty in automatically ranking examples from easy to hard and determining optimal pacing for introducing harder samples.
• Despite its benefits, improper implementation of curriculum learning can lead to convergence issues, as models may get trapped in suboptimal local minima.
• Vision transformers and medical imaging datasets present promising but underexplored application areas for curriculum learning, especially when using data-level curricula based on image complexity or labeling quality.
• Current curriculum strategies in NLP often rely on simplistic heuristics like text length, suggesting a need for more human-like curricula based on vocabulary complexity and linguistic structure.
• Future integration of curriculum learning with alternative optimization algorithms—such as evolutionary search—may close the performance gap between standard SGD and more advanced optimizers.