# Curriculum Learning: A Survey

**Authors:** Petru Soviany, Radu Tudor Ionescu, Paolo Rota, Nicu Sebe  
**arXiv:** [2101.10382](https://arxiv.org/abs/2101.10382)  
**Date:** 2021-01-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2101.10382v1/

## Key points

- Curriculum learning involves training models in a specific order, starting from easy samples and gradually progressing to harder ones
- This approach improves performance compared to randomly shuffling data during training
- Curriculum learning has been successfully applied across various areas of machine learning and tasks
- Challenges include ranking samples from easy to hard and determining the appropriate pacing function for introducing more difficult data
- The survey explores existing literature and discusses how researchers have tackled these challenges
- Different instantiations of curriculum learning for various machine learning tasks are presented, providing insights into their effectiveness
- A multi-perspective taxonomy of curriculum learning approaches is constructed, considering various classification criteria
- A hierarchical tree of curriculum learning methods is developed using an agglomerative clustering algorithm, linking clusters with the taxonomy
- The survey offers valuable insights into curriculum learning and its potential applications in machine learning
- Areas for future research and development in this field are identified

## Summary

The paper titled "Curriculum Learning: A Survey" explores the concept of curriculum learning and its impact on training machine learning models. Curriculum learning involves training models in a specific order, starting from easy samples and gradually progressing to harder ones. This approach has been found to improve performance compared to the standard training method that randomly shuffles data, without incurring additional computational costs. The authors highlight that curriculum learning strategies have been successfully applied across various areas of machine learning and a wide range of tasks. However, one challenge lies in determining how to rank the samples from easy to hard and identifying the appropriate pacing function for introducing more difficult data. These factors can limit the practical application of curriculum approaches. To address these limitations, the survey delves into existing literature and discusses how researchers have tackled these challenges. The authors present different instantiations of curriculum learning for various machine learning tasks, providing insights into their effectiveness. In order to provide a comprehensive overview, the authors construct a multi-perspective taxonomy of curriculum learning approaches by considering various classification criteria. They also develop a hierarchical tree of curriculum learning methods using an agglomerative clustering algorithm, linking the discovered clusters with their taxonomy. Overall, this survey offers valuable insights into curriculum learning and its potential applications in machine learning. It not only presents existing approaches but also identifies areas for future research and development in this field; thus providing guidance on how best to leverage this technique for improved performance in machine learning tasks.