Słownik AI
Kompletny słownik sztucznej inteligencji
Discriminative Learning Rate
Fine-tuning strategy applying different learning rates to network layers, typically a lower rate for pre-trained layers and a higher rate for new layers.
Layer Freezing
Technique involving disabling weight updates (gradients) of certain layers in the pre-trained model during training to preserve learned generic features.
Detection Head
Module added at the end of the backbone network, specifically trained to predict bounding boxes and probability classes for detected objects.
Progressive Retraining
Fine-tuning approach where the model is first trained with upper layers frozen, then lower layers are progressively unfrozen for finer adjustment.
ImageNet Weights
Pre-trained parameters on the ImageNet dataset (1.2 million images, 1000 classes), commonly used as a starting point for computer vision models in transfer learning.
Domain Bias
Systematic difference in the statistical distribution of data between the source domain (where the model is pre-trained) and the target domain (where it is deployed).