Meta-Learning

Tasks, adaptation and learned components

Learning across tasks — 2020. 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.
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.

MAML — 2017. 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.
First-order MAML and Reptile — 2018. 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.
Meta-SGD — 2017. 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.
Implicit MAML — 2019. 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.
ANIL and feature reuse — 2019. 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.

Matching Networks — 2016. 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.
Prototypical Networks — 2017. 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.
Relation Networks — 2017. 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.
MetaOptNet — 2019. Learns embeddings through differentiable convex task classifiers, including linear SVMs. Implicit differentiation of optimality conditions links the fitted task head to representation learning.

Learned optimizers — 2016. 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.
VeLO — 2022. 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.
ALPaCA — 2018. 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.
Hierarchical Bayes view of MAML — 2018. 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.
PACOH — 2021. 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.

Conditional Neural Processes — 2018. 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.
Prior-Data Fitted Networks — 2021. 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.

RL² — 2016. 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.
PEARL — 2019. 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.
Meta-World — 2019/2020. 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.

Online meta-learning / FTML — 2019. 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.

SimpleShot — 2019. 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.
A Closer Look at Few-shot Classification — 2019. 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.
Meta-Dataset — 2019. 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.
Cross-domain few-shot evidence — 2019. 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.
Meta-learning without memorization — 2019. 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.

learn2learn — 2020. Provides reusable meta-learning routines, task datasets and algorithm interfaces. Standardized components help reproduce adaptation experiments; they do not ensure equivalent task splits or tuning budgets.

Star Issue