Meta-Learning: A Survey

By Joaquin Vanschoren. 2018

844 citations recorded by the source corpus.

Meta-learning, or learning to learn, is the science of systematically observing how different machine learning approaches perform on a wide range of learning tasks, and then learning from this experience, or meta-data, to learn new tasks much faster than otherwise possible. Not only does this dramatically speed up and improve the design of machine learning pipelines or neural architectures, it also allows us to replace hand-engineered algorithms with novel approaches learned in a data-driven way. In this chapter, we provide an overview of the state of the art in this fascinating and continuously evolving field.

Subfields: Artificial Intelligence, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Representation & Self-Supervised Learning

Citations in the corpus

Finetuned Language Models Are Zero-Shot Learners (2021, background): “In addition, meta-learning (Finn et al., 2017; Vanschoren, 2018, inter alia) also broadly tries to train models that adapt quickly to unseen tasks, typically based on a few examples.”
Differentially Private Meta-Learning (2019, background): “Note that QG is significantly weaker than strong convexity, which previous work (Finn et al., 2019) has assumed to hold for task losses but does not hold for applicable cases such as few-shot least-squares or logistic regression if the number of task-samples is smaller than the data-dimension.”
Hyper-Parameter Optimization: A Review of Algorithms and Applications (2020, neutral): “The target model is supposed to have a similar structure to the source model (Vanschoren, 2018; Yosinski et al., 2014); thus, transfer learning is especially feasible for hyper-parameter tuning.”
Automated Machine Learning: State-of-The-Art and Open Challenges (2019, background): “These techniques can generally be categorized into three broad groups [123]: learning based on task properties, learning from previous model evaluations and learning from already pretrained models (Figure 3).”
Survey on Automated Machine Learning (2019, methodology): “ or measures from information theory. Using the meta-features of a new task m~(t new), a subset of P0ˆP with similar tasks can be obtained. P0is than used similarly to task-independent meta-learning (Vanschoren, 2018). 8. Selected Frameworks This section provides an introduction to the evaluated Aut”
CoMPS: Continual Meta Policy Search (2021, methodology): “This differs from the standard reinforcement learning setup, in which there is a single task T0, and the standard meta-RL setting, where all tasks can be revisited as many times as needed during meta-training but is closer to and RL approximation of online meta learning (Finn et al., 2019).”
Continual Learning via Local Module Composition (2021, background): “Continual-meta learning focuses on fast learning and remembering [25, 35, 33, 41], often emphasising the online performance on OOD tasks [14].”
Towards Continual Reinforcement Learning: A Review and Perspectives (2020, background): “There have been a number of recent improvements on this topic of meta-optimization in multi-task RL (Nichol and Schulman, 2018; Kim et al., 2018; Rothfuss et al., 2018; Flennerhag et al., 2018; Nagabandi et al., 2018; Mendonca et al., 2019; Finn et al., 2019; Lin et al., 2020; Berseth et al., 2021; ”
Incremental Meta-Learning via Indirect Discriminant Alignment (2020, methodology): “In IML, we want to continuously improve the meta-trained model so that, presented with an unseen task, it achieves better performance now that it would have before solving the previous task.”
Meta-Learning in Neural Networks: A Survey (2020, methodology): “Some early steps were made towards deﬁning meta-learning ready continual benchmarks in [96], [160], [253], mainly composed of Omniglot and perturbed versions of MNIST.”
Few-Shot Learning via Learning the Representation, Provably (2020, background): “Another recent line of theoretical work analyzed gradient-based meta-learning methods (Denevi et al., 2019, Finn et al., 2019, Khodak et al., 2019) and showed guarantees for convex losses by using tools from online convex optimization.”
La-MAML: Look-ahead Meta Learning for Continual Learning (2020, background): “[2, 10, 19] investigate orthogonal setups in which a learning agent uses all previously seen data to adapt quickly to an incoming stream of data, thereby ignoring the problem of catastrophic forgetting.”
Unsupervised Progressive Learning and the STAM Architecture (2019, methodology): “Few-shot learning (FSL) and Meta-learning: These methods recognize object classes not seen in the training set with only a single (or handful) of labeled examples [Vanschoren, 2018].”
Theory of Mind and Preference Learning at the Interface of Cognitive Science, Neuroscience, and AI: A Review (2022, methodology): “…situations in which we have access to data from multiple overlapping tasks, however, we can employ techniques from multi-task learning (Caruana, 1997), meta-learning (Vanschoren, 2018), and/or continual learning (Van de Ven and Tolias, 2019) to learn from the available training data more eﬃciently.”