AI-woordenlijst
Het complete woordenboek van kunstmatige intelligentie
Genetic Algorithms for NAS
Optimization methods inspired by biological evolution using selection, crossover, and mutation to automatically discover efficient network architectures.
Evolution Strategies
Evolutionary optimization algorithm where mutation is the primary variation operator, adapted to efficiently explore the neural architecture space.
Architecture Search Space
Set of all possible neural network architectures defined by structural and operational constraints specific to the problem.
NAS Fitness Function
Multi-criteria evaluation function measuring the performance of a neural architecture based on accuracy, computational complexity, and memory consumption.
NAS Computation Cells
Repeatable modular structures discovered by NAS, combining different convolution and connection operations to build complex architectures.
Architecture Crossover
Genetic operator combining substructures of two parent architectures to create potentially performant new architectures.
Architecture Mutation
Stochastic modification of an existing neural architecture by adding, removing, or modifying layers to explore new configurations.
Differential Evolution NAS
Evolutionary algorithm variant using differential vectors to guide mutation in the continuous space of architecture hyperparameters.
Multi-objective NAS
NAS approach that simultaneously optimizes multiple conflicting objectives such as accuracy, latency, and energy consumption through Pareto methods.
Model sampling
Probabilistic selection technique of candidate architectures in the evolutionary population for evaluation and reproduction based on their fitness.
NAS primitive operations
Basic set of neural operations (convolution, pooling, activation) used as fundamental building blocks to construct complex architectures.
Elitism in NAS
Strategy that systematically preserves the best architectures found at each generation to ensure monotonic performance convergence.
Evolutionary AutoML
Complete machine learning system primarily using evolutionary methods to fully automate the ML pipeline including NAS.
Pareto path NAS
Exploration of the Pareto front in the multi-objective space to identify a set of architectures offering different performance-cost trade-offs.
Neuro-evolutionary search
Interdisciplinary field combining computational neuroscience and evolutionary algorithms to discover bio-inspired architectures.
Architecture hyperparameters
Structural parameters of a network (depth, width, connection types) optimized by NAS beyond classical training hyperparameters.
Progressive NAS Evaluation
Accelerated evaluation strategy using proxies or data subsets to quickly estimate the fitness of candidate architectures.