YZ Sözlüğü
Yapay Zekanın tam sözlüğü
Neural Architecture Search (NAS)
Process of automating the design of optimal neural network architectures for a specific task by exploring a defined search space.
Reinforcement Learning based NAS
NAS approach using a reinforcement learning agent that generates network architectures and receives rewards based on their performance.
Evolutionary NAS
NAS method inspired by biological evolution where architectures are selected, crossed over, and mutated to progressively improve their performance.
Gradient-based NAS
NAS technique that optimizes the architecture using gradients, enabling a continuous and differentiable search in the architecture space.
Cell-based NAS
NAS strategy where the search space is decomposed into reusable computational cells, reducing the complexity of global architecture search.
Search Space
Set of all possible neural network architectures that NAS can explore, defined by structural and operational constraints.
Performance Estimation Strategy
Method for quickly evaluating the performance of candidate architectures without full training, crucial for the computational efficiency of NAS.
Differentiable Architecture Search (DARTS)
Revolutionary NAS algorithm that relaxes discrete search into a continuous optimization problem, allowing architecture discovery through gradient descent.
Network Morphism
A network transformation technique that preserves functionality while modifying the structure, used to efficiently explore the architecture space.
Meta-learning in NAS
The application of meta-learning to accelerate NAS by learning from previous searches on different tasks or datasets.
Hardware-aware NAS
An NAS approach that incorporates specific hardware constraints and characteristics into the search process to optimize performance on the target.
Zero-cost NAS Proxies
Ultra-fast evaluation metrics that estimate the performance of architectures without training, based on gradient or activation statistics.
Neural Tangent Kernel in NAS
The use of the NTK to predict network performance without full training, serving as an effective proxy in modern NAS strategies.