Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
Supernetwork
Over-parameterized neural network containing all candidate architectures as subgraphs, enabling simultaneous evaluation of multiple architectures through weight sharing between different configurations.
Path Sampling
Random sampling method of subpaths in the supernetwork during training, ensuring all potential architectures receive equal attention and avoiding selection bias.
Sub-network Sampling
Process of extracting and evaluating specific architectures from the trained supernetwork, sampling different configurations to identify the best-performing ones without additional training cost.
Performance Predictor
Model trained to estimate the performance of an architecture without full evaluation, using structural features to efficiently guide the search toward promising architectures.
Once-for-All Network
Advanced One-Shot NAS architecture enabling deployment on different hardware constraints without retraining, using a progressively shrunk supernetwork to adapt to available resources.
Proxyless NAS
One-Shot NAS approach eliminating the use of proxy datasets or smaller models, training directly on the target dataset while maintaining computational efficiency.
Single-path NAS
One-Shot NAS variant simplifying the supernetwork to a single path during training, reducing complexity and improving stability while preserving architectural exploration capability.
FairNAS
One-Shot NAS approach ensuring fair treatment of all candidate architectures during training, eliminating performance biases and ensuring impartial evaluation of configurations.
FBNet
Facebook Berkeley Network, One-Shot NAS framework that directly integrates hardware constraints into the search process, optimizing architectures for specific platforms in real time.
Neural Architecture Parameters
Continuous parameters representing the weights of architectural decisions in differentiable approaches, optimized alternately with network weights to converge towards an optimal architecture.
Architecture Gradient
Gradient of architecture parameters with respect to the objective function, allowing direct update of structural decisions in One-Shot NAS approaches based on differentiable optimization.