Słownik AI
Kompletny słownik sztucznej inteligencji
Parameter Initialization
Crucial process in MAML that involves finding optimal starting weights that minimize the adaptation distance needed to achieve good performance on a new task.
Support Tasks
Subset of data from a meta-training task used to compute adaptation gradients and temporarily update the model parameters.
Query Tasks
Validation samples in each task used to evaluate performance after adaptation and compute meta-gradients for updating initialization parameters.
Meta-learning Step
External learning rate used in the meta-optimization loop to update initialization parameters by minimizing loss on query sets.
Task Adaptation Step
Internal learning rate applied during rapid adaptation to a new specific task in MAML's internal optimization loop.
Meta-gradients
Gradients calculated through task adaptation steps, allowing performance information to propagate to initialization parameters.
First-Order MAML (FOMAML)
Computationally efficient variant of MAML that ignores second derivatives in meta-gradients, reducing complexity while maintaining good performance.
Inner Loop
Internal optimization loop in MAML that performs rapid adaptation of parameters to a specific task using support data.
Outer loop
Outer optimization loop in MAML that updates initialization parameters by aggregating performance information across all tasks.
Task distribution
Underlying set of tasks from which MAML samples during training to learn robust and generalizable representations.
Reptile
Simplified meta-learning algorithm that performs interpolation between initialized weights and weights after adaptation, without requiring nested gradients.
Cross-task generalization
Fundamental objective of MAML consisting of learning representations that effectively transfer knowledge between different related tasks.
Task uncertainty quantification
Extension of MAML incorporating Bayesian methods to quantify uncertainty in predictions when adapting to new tasks.