KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
ResNet
Residual neural network architecture introduced by Microsoft in 2015, using shortcut connections to enable training of very deep networks without performance degradation.
VGG
Convolutional neural network architecture developed by Oxford University, characterized by its exclusive use of stacked 3x3 filters and variable depth (VGG16, VGG19).
Convolutional Neural Network
Type of deep neural network architecture specially designed to process grid-like data such as images, using convolution operations to extract hierarchical features.
ImageNet
Reference database containing over 14 million labeled images in 1000 categories, commonly used to pre-train computer vision models.
Model Freezing
Technique consisting of disabling weight updates for certain network layers during training, typically applied to lower layers of a pre-trained model.
Layer-wise Learning Rates
Optimization strategy applying different learning rates according to network layers, generally higher for upper layers and lower for lower pre-trained layers.
Head Replacement
Transfer learning technique where the final classification layer of a pre-trained model is replaced by a new layer adapted to the number of classes in the target task.
Gradual Unfreezing
Fine-tuning strategy where network layers are gradually unfrozen, starting with upper layers and descending towards lower layers over epochs.