Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Reinforcement Learning Controller
Reinforcement learning agent that generates neural network architectures in sequence, optimizing its decisions based on rewards obtained from the performance of evaluated architectures.
Monte Carlo Tree Search for NAS
Exploration algorithm that uses random simulations to evaluate and select promising architectures in the neural network search space.
RNN Controller
Recurrent neural network used as a controller in reinforcement learning NAS, sequentially generating architecture descriptions in the form of tokens or actions.
Search Space Pruning
Technique for reducing the search space by eliminating suboptimal architectures based on heuristics or preliminary evaluations to accelerate the NAS process.
Meta-Learning for NAS
Approach where the system learns to learn to design architectures, transferring knowledge from previous tasks to accelerate search on new tasks.
Weight Sharing Strategy
Technique where the weights of architectures are shared during training, allowing rapid evaluation of multiple architectures without resetting parameters.
Reward Function Design
Mathematical definition of the reward in reinforcement learning NAS, typically combining model accuracy, computational complexity, and other constraints.
Latency-aware NAS
Variant of NAS that explicitly optimizes architectures to minimize inference latency in addition to accuracy, crucial for real-time applications.
Architecture Regularization
Application of regularization techniques directly on the architecture space to avoid overfitting and promote generalizable solutions.
Early Stopping in NAS
Early stopping mechanism for the evaluation of promising architectures based on early performance indicators to optimize computational budget.
Neural Tangent Kernel Analysis
Theoretical tool used in NAS to analyze convergence and expressivity properties of candidate architectures before their full training.