Source-linked AI summary
Curriculum Learning: A Survey
Petru Soviany, Radu Tudor Ionescu, Paolo Rota, Nicu Sebe
TL;DR
Curriculum learning can improve machine-learning training by ordering examples from easy to hard, but ranking samples and scheduling difficulty remain limiting challenges. This survey formalizes and categorizes curriculum methods across machine-learning components and corroborates its taxonomy with hierarchical clustering, finding successful applications across domains and tasks despite inconsistent gains.
Problem
Curriculum learning requires reliable sample-ranking and difficulty-scheduling strategies, which can limit its broader use despite reported benefits over random data ordering.
Method
The survey formalizes curriculum learning, builds a multi-perspective taxonomy, and uses Ward-linkage agglomerative clustering to construct a hierarchical tree of methods.
Results
Curriculum learning has improved performance across diverse machine-learning domains and tasks, while the automatically clustered method tree corroborates the manually constructed taxonomy.
Takeaways & Limitations
The survey provides a unified framework for comparing curriculum-learning approaches across data, models, tasks, performance measures, ranking criteria, and schedules.
Takeaways & Limitations
Curriculum learning does not always significantly improve performance and may degrade data diversity, with outcomes influenced by other potentially negative factors.
Abstract
from arXiv · showhide
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.
1 Introduction
The introduction motivates curriculum learning as an alternative to randomly ordered training, inspired by how humans progress from easy concepts to hard ones. It presents a unified formulation, taxonomy, and analysis linking curriculum learning to machine-learning components and loss smoothing.
- Context and motivation: Deep neural networks achieve state-of-the-art results across diverse tasks, but comparatively less attention has focused on improving their training process.
- Context and motivation: Standard neural-network training usually presents randomly selected mini-batches, so examples are considered in a random order.
- Context and motivation: Curriculum learning is motivated by human learning, in which basic easy concepts are learned before advanced hard concepts rather than being presented randomly.
- Contributions: The survey formalizes existing curriculum learning methods under one umbrella and defines a generic formulation.
- Contributions: It links curriculum learning to data, model, task, and performance measure components, interprets these forms through loss-function smoothing, and constructs a manual taxonomy.
- Related surveys and organization: The survey also situates its analysis alongside curriculum-learning surveys in reinforcement learning and outlines sections covering formulation, taxonomy, applications, clustering, and future directions.
2 Curriculum Learning
Curriculum learning gradually increases training complexity, aiming to improve model performance faster through smoother objectives. The survey generalizes this idea across data, model, and performance-measure levels using criteria and schedulers that control curriculum progression.
- 2 Curriculum Learning: Curriculum learning gradually increases data-sample complexity during training, imitating human learning and aiming to improve performance faster.Using easy samples initially is expected to make the objective function smoother; increasing difficulty makes it more complex.
- 2 Curriculum Learning: Data-level and model-level curricula share a scheduler and performance measure, with the scheduler controlling when curriculum updates occur and which pace is used.Possible pacing functions include linear and logarithmic schedules.
- 2 Curriculum Learning: Continuation methods can be interpreted as curriculum learning over the performance measure, although they predate curriculum learning and remain an independent research field.This connection is not typically mentioned in the literature.
- 2 Curriculum Learning: The generic curriculum formulation applies a curriculum criterion at a chosen level, such as data, model, or performance measure, while a scheduler governs updates during training.The criterion determines ordering or difficulty, and training repeatedly selects data and updates the model.
- 2 Curriculum Learning: Difficulty criteria can be task-dependent, including shape complexity for images, grammar properties for text, and signal-to-noise ratio for audio.The survey also notes more general curriculum-generation methods beyond task-specific criteria.
- 2 Curriculum Learning: Easy-to-hard ordering can be generated by teacher networks, model learning progress, or task sequencing when multiple tasks are involved.These correspond to teacher-student supervision, self-paced learning, and ordering tasks to maximize the final result.
3 Taxonomy of Curriculum Learning Methods
The survey proposes a balanced, seven-category taxonomy of curriculum learning methods because existing component-based categorizations are heavily dominated by data-level curricula. It also distinguishes approaches by curriculum-building criteria, scheduling strategies, and application level.
- Taxonomy overview: The survey replaces an unbalanced component-based categorization, dominated by data-level methods, with a more balanced partition of seven categories.The categories arise from different assumptions and model requirements.
- Methodology-based categories: Vanilla curriculum learning uses a priori rules to select samples in an increasingly difficult order.Bengio et al. introduced Vanilla CL in 2009 and showed performance improvements when models receive progressively harder samples.
- Methodology-based categories: Self-paced learning computes sample ordering from the model’s own performance rather than specifying it a priori, allowing the order to vary during training.Examples rank samples using prediction likelihood or objectness.
- Methodology-based categories: Balanced curriculum learning adds sample diversity to traditional difficulty-based ordering, introducing multiple ordering criteria.Its difficulty criterion still presents easy samples first and progressively introduces harder ones.
- Other curriculum dimensions: Curriculum methods also include teacher-guided, implicit, anti-curriculum, and uncertainty-based active-learning variants, alongside different curriculum criteria, schedules, and application levels.Anti-CL can order examples from hard to easy using signal-to-noise ratio, while active learning selects by uncertainty rather than difficulty.
4 Applications of Curriculum Learning
The section surveys curriculum learning applications across domains and tasks, covering strategies that order examples by difficulty, diversify sampling, adapt schedules, or structure task progression. Reported applications span speech, vision, language, reinforcement learning, face recognition, generation, and dialogue-related tasks.
- Section organization: Applications are organized first by domain and then by task, presenting similar approaches together to trace curriculum methodology development.The section’s ordering is intended to help readers locate work relevant to their field of interest.
- Computer vision: Face recognition curricula use head pose, including yaw, pitch, and roll, to present upright frontal faces before harder examples and improve the random baseline.Head pose serves as the difficulty measure, with upright frontal faces treated as easiest to recognize.
- Computer vision: Self-paced learning with diversity combines easy-first selection with sufficiently diverse examples to reduce bias toward easy samples from the same group.The cited methodology extends standard easy-to-hard approaches by jointly considering sample difficulty and diversity.
- Natural language processing: In natural language generation, curricula increase sequence length gradually, and curriculum ordering is essential for generating long LSTM sequences.Adversarial architectures constrain generators to produce sequences of progressively increasing lengths.
- Natural language processing: Data-to-text experiments find that joint data-and-text difficulty measures outperform measures capturing only data complexity or text complexity.The curriculum selects examples easy enough for the model’s competence at the current training step.
- Other applications: Curriculum methods improve training across intent detection, dialogue policy learning, cross-domain translation, image classification, and masked-goal reinforcement learning applications.These methods include dynamic difficulty scheduling, teacher-student difficulty-and-diversity control, meta-learning curricula, easy-to-hard optimization, and goal masking.
5 Hierarchical Clustering of Curriculum Learning Methods
The survey supplements its manually constructed taxonomy with Ward-linkage agglomerative clustering, whose dendrogram reveals homogeneous curriculum-learning groups with few outliers. The resulting hierarchy first separates learning paradigms and then subdivides supervised methods by application domain, consistent with the manual taxonomy.
- Clustering methodology: The authors use Ward-linkage agglomerative clustering to automate grouping and reduce potential subjectivity in the manually constructed taxonomy.They chose hierarchical clustering because it performs well despite noise between clusters and suits the fine-grained grouping problem.
- Clustering results: The dendrogram produces homogeneous clusters containing at most one or two outlier papers.Figure 2 presents the resulting hierarchical grouping of curriculum-learning articles.
- Clustering results: Reinforcement-learning curricula form a distinct cluster because they typically use teacher-student models or operate over tasks, unlike data-sample curricula in other domains.The reinforcement-learning group is identified as the green cluster, while other domains occupy the brown, blue, purple, and red clusters.
- Hierarchical taxonomy: The dendrogram first divides curriculum-learning studies by supervised learning, reinforcement learning, and self-paced learning.This paradigm-level division is presented as the primary organizational structure suggested by the hierarchical clustering.
- Hierarchical taxonomy: Supervised-learning methods are further divided by application area into image and text processing, speech processing, object detection and segmentation, and domain adaptation.The authors conclude that this second-level structure is consistent with their manually determined taxonomy.
6 Closing Remarks and Future Directions
The survey concludes that curriculum learning can improve performance across many machine-learning domains, but its success depends on difficulty measures and remains insufficiently understood in relation to SGD. It identifies underexplored directions including unsupervised and self-supervised learning, transformers, and domain-specific applications.
- Closing Remarks: Curriculum learning can degrade data diversity and produce worse results, while its success is strongly correlated with the difficulty measure used to rank samples or tasks.Difficulty measures that incorporate diversity appear to yield higher improvements than measures that overlook sample diversity.
- Future Directions: The survey finds a deficit of curriculum learning studies in unsupervised learning, particularly self-supervised learning.Existing strategies have been investigated with supervised, cross-domain adaptation, self-paced, semi-supervised, and reinforcement learning paradigms.
- Closing Remarks: The connection between curriculum learning and stochastic gradient descent remains insufficiently understood because sample order affects stochastic training in neural networks with non-convex objectives.Curriculum learning is typically applied to neural networks, whose non-convex models are commonly optimized using SGD variants.
- Curriculum learning in computer vision: Curriculum learning has not been applied to vision transformers, motivating data-level and model-level curricula during fine-tuning.Suggested approaches include an image-difficulty predictor attentive to data diversity and gradually unsmoothing tokens to seek accuracy and efficiency gains.
- Curriculum learning in medical imaging: Medical image transformers represent a promising future application because curriculum learning has not yet been studied with them.The survey frames this direction as following the emerging transformer trend in computer vision and medical imaging.
- Curriculum learning in natural language processing: Future curriculum strategies in natural language processing should move beyond simple heuristics, while signal-processing curricula could organize samples by domain-specific properties such as noise level.Language-transformer curricula have already been explored, whereas signal-processing challenges include denoising and source separation.