# Meta-Learning ## Tasks, adaptation and learned components - **Learning across tasks** - Meta-learning optimizes a learning procedure using experience from multiple tasks or learning episodes. Hospedales et al. distinguish what is learned, the outer objective and the adaptation mechanism; few-shot learning is one application. [Hospedales et al. 2021]( "Meta-Learning in Neural Networks: A Survey; IEEE Transactions on Pattern Analysis and Machine Intelligence"). - **Task-level evaluation** - Separate meta-training, validation and test tasks, with support data for adaptation and query data for assessment. Unseen classes and unseen domains are different tests; report both the task distribution and adaptation budget. [Triantafillou et al. 2020]( "Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples; ICLR"). ## Learning initializations and adaptation rules - **MAML** - Optimizes initialization so a few gradient steps on a new task yield useful predictions. The outer objective evaluates the adapted model; this differs from ordinary pooled pretraining followed by an unoptimized adaptation rule. [Finn et al. 2017]( "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks; ICML"). - **Meta-SGD** - Learns an initialization and per-parameter update coefficients for one-step adaptation. The coefficients determine update direction and scale, increasing the learned adaptation rule's flexibility beyond a shared learning rate. [Li et al. 2017]( "Meta-SGD: Learning to Learn Quickly for Few-Shot Learning; arXiv"). - **First-order MAML and Reptile** - Avoid second-order derivatives in meta-updates. Reptile repeatedly trains on sampled tasks and moves the shared initialization toward the resulting parameters; it does not use the same update as full MAML. [Nichol et al. 2018]( "On First-Order Meta-Learning Algorithms; arXiv"). - **Hierarchical Bayes view of MAML** - Grant et al. interpret shared initialization and task adaptation as approximate hierarchical Bayesian inference. This connects learned inductive bias to a prior across tasks; gradient adaptation does not compute an exact posterior. [Grant et al. 2018]( "Recasting Gradient-Based Meta-Learning as Hierarchical Bayes; arXiv"). - **Implicit MAML** - Computes meta-gradients through the inner problem's solution rather than storing the optimization path. Rajeswaran et al. use implicit differentiation; gradient accuracy depends on solving the relevant inner and linear systems. [Rajeswaran et al. 2019]( "Meta-Learning with Implicit Gradients; NeurIPS"). - **ANIL and feature reuse** - Adapts only the task-specific head in the inner loop. Raghu et al. match MAML on studied few-shot image and RL benchmarks, finding feature reuse dominant there; this does not show adaptation is never useful. [Raghu et al. 2020]( "Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAML; ICLR"). ## Metric-based and optimization-based task heads - **Matching Networks** - Predicts query labels by attention over a labeled support set in learned feature space. Episodes train the support-to-query mapping; new classes can be handled without gradient fine-tuning. [Vinyals et al. 2016]( "Matching Networks for One Shot Learning; NeurIPS"). - **Prototypical Networks** - Represents each support class by its mean embedding and classifies queries by distance to these prototypes. A simple task head makes embedding quality and distance choice central. [Snell et al. 2017]( "Prototypical Networks for Few-shot Learning; NeurIPS"). - **Relation Networks** - Learns a comparison function between query and support representations within episodes. Unlike a fixed Euclidean metric, the similarity function is trained; novel-class predictions require no additional weight updates. [Sung et al. 2018]( "Learning to Compare: Relation Network for Few-Shot Learning; CVPR"). - **MetaOptNet** - Learns embeddings through differentiable convex task classifiers, including linear SVMs. Implicit differentiation of optimality conditions links the fitted task head to representation learning. [Lee et al. 2019]( "Meta-Learning with Differentiable Convex Optimization; CVPR"). ## Learned optimizers and Bayesian adaptation - **Learned optimizers** - Andrychowicz et al. train an LSTM to produce optimization updates. Their results transfer to tasks with similar structure; optimizer quality depends on the training distribution and optimization horizon. [Andrychowicz et al. 2016]( "Learning to learn by gradient descent by gradient descent; NeurIPS"). - **ALPaCA** - Meta-learns neural features and a prior over linear weights from sample functions. Bayesian linear regression then adapts online in feature space; posterior uncertainty is conditional on the learned model. [Harrison et al. 2020]( "Meta-Learning Priors for Efficient Online Bayesian Regression; Springer Proceedings in Advanced Robotics Algorithmic Foundations of Robotics XIII"). - **PACOH** - Uses PAC-Bayesian bounds to learn a distribution over priors and control meta-overfitting. The analysis assumes independently sampled tasks from a common environment and independently sampled data within each task. [Rothfuss et al. 2021]( "PACOH: Bayes-Optimal Meta-Learning with PAC-Guarantees; ICML"). - **VeLO** - Scales optimizer meta-training across many tasks. The paper reports transfer gains but also failures on much larger models and longer training runs outside its training distribution; optimizer reuse needs evaluation across both axes. [Metz et al. 2022]( "VeLO: Training Versatile Learned Optimizers by Scaling Up; arXiv"). ## Adaptation through context - **Conditional Neural Processes** - Aggregate observed input/output pairs into a permutation-invariant context representation and predict at new inputs. Adaptation changes the context rather than network weights; the network is trained across functions. [Garnelo et al. 2018]( "Conditional Neural Processes; ICML"). - **Prior-Data Fitted Networks** - Train on tasks sampled from a chosen prior to approximate posterior predictions. A new dataset becomes context for a forward pass; the learned inference procedure inherits the synthetic task distribution, not a universal Bayesian prior. [Müller et al. 2022]( "Transformers Can Do Bayesian Inference; ICLR"). ## Meta-reinforcement learning - **RL²** - Trains a recurrent policy across tasks; hidden state carries experience between episodes within a task. Fast adaptation occurs in activations while the recurrent weights encode the learned learning procedure. [Duan et al. 2016]( "RL$^2$: Fast Reinforcement Learning via Slow Reinforcement Learning; arXiv"). - **PEARL** - Separates task inference from control using probabilistic latent context and off-policy RL. Posterior sampling supports exploration; adaptation infers the task from experience rather than retraining the whole policy. [Rakelly et al. 2019]( "Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables; ICML"). - **Meta-World** - Fifty simulated manipulation tasks test learning new behaviors rather than only new goal parameters. In the original study, the tested methods struggle with joint learning even across ten tasks that are individually learnable. [Yu et al. 2020]( "Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning; CoRL"). ## Sequential task learning - **Online meta-learning / FTML** - Updates an initialization as tasks arrive, evaluating performance after adaptation. Finn et al. give logarithmic regret under strong-convexity and smoothness conditions; their neural experiments extend beyond that theorem's setting. [Finn et al. 2019]( "Online Meta-Learning; ICML"). ## Baselines and comparative findings - **SimpleShot** - Mean-centering and normalizing pretrained features make nearest-neighbor classification competitive on studied few-shot benchmarks. Wang et al. show that gains need comparison with strong feature-based baselines without episodic meta-training. [Wang et al. 2019]( "SimpleShot: Revisiting Nearest-Neighbor Classification for Few-Shot Learning; arXiv"). - **A Closer Look at Few-shot Classification** - Chen et al. find deeper backbones shrink method differences on related domains, while ordinary fine-tuning is competitive under domain shift. Architecture and adaptation protocol can dominate the algorithm label. [Chen et al. 2019]( "A Closer Look at Few-shot Classification; ICLR"). - **Meta-Dataset** - Evaluates varying few-shot tasks drawn from diverse datasets rather than one fixed benchmark setup. Triantafillou et al. analyze how task characteristics and training-source diversity affect generalization. [Triantafillou et al. 2020]( "Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples; ICLR"). - **Cross-domain few-shot evidence** - Guo et al. evaluate natural, satellite and medical images; simple fine-tuning outperforms the tested meta-learning methods on average. The finding concerns substantial domain shifts in this benchmark. [Guo et al. 2020]( "A Broader Study of Cross-Domain Few-Shot Learning; ECCV"). - **Meta-learning without memorization** - A meta-learner can solve training tasks while ignoring support data. Yin et al. use information-theoretic regularization to encourage adaptation on non-mutually-exclusive tasks; low meta-training loss alone can hide this shortcut. [Yin et al. 2020]( "Meta-Learning without Memorization; ICLR"). ## Software - **learn2learn** - Provides reusable meta-learning routines, task datasets and algorithm interfaces. [Arnold et al. 2020]( "learn2learn: A Library for Meta-Learning Research; arXiv"). ## Related topics - [Transfer learning](https://mlpapers.org/transfer-learning/) - [AutoML](https://mlpapers.org/automl/) - [Reinforcement learning](https://mlpapers.org/reinforcement-learning/) - [Federated learning](https://mlpapers.org/federated-learning/) - [Generalization](https://mlpapers.org/generalization/) - [Evaluation](https://mlpapers.org/evaluation/)