AI-ordlista
Den kompletta ordlistan över AI
Population-Based Training (PBT)
Hyperparameter optimization method that combines reinforcement learning and evolutionary algorithms to dynamically adjust hyperparameters during model training.
Adaptive Hyperparameters
Training parameters that evolve dynamically based on model performance rather than being pre-fixed.
Population Dynamics
Evolutionary process where individuals (models) interact, compete, and exchange information in a shared training environment.
Dynamic Learning Schedule
Automatic learning rate adjustment strategy based on relative performance within the population.
Evaluation Period
Time interval between selection and exploitation phases in the PBT cycle, determining the frequency of hyperparameter adjustments.
Performance Metric
Metric or set of metrics used to evaluate and compare models within the PBT population.
Adaptive Exponential Decay
Automatic learning rate adjustment strategy where decay is dynamically modulated based on relative performance.
Inter-individual Knowledge Transfer
PBT mechanism allowing models to learn collectively by sharing the best configurations discovered within the population.
Nash equilibrium in PBT
Theoretical state where no model in the population can improve its performance by unilaterally modifying its hyperparameters.
PBT multi-objective optimization
Extension of PBT that simultaneously handles multiple conflicting objectives such as accuracy vs inference time or energy consumption.
Evolutionary regularization
Beneficial side effect of PBT where competition and population diversity act as an implicit regularization mechanism.
Guided stochastic exploration
PBT exploration strategy where random mutations are biased towards regions of the hyperparameter space that have shown potential.