AI Glossary
The complete dictionary of Artificial Intelligence
Deep Convolutional Layers
Initial layers of a convolutional neural network that capture low-level features like edges, textures, and fundamental geometric shapes. These layers are typically reused as-is during transfer learning due to their ability to extract universal patterns.
Pre-trained Neural Networks
Deep learning models already trained on large datasets like ImageNet, possessing optimized weights for generic feature extraction. These networks serve as a foundation for transfer learning by providing powerful feature extractors.
Feature Vectors
Multidimensional numerical representations produced by the intermediate layers of a pre-trained neural network. These vectors encode essential semantic information from input data in a compact and structured space.
Partial Fine-Tuning
Transfer learning strategy where only the upper layers of the model are retrained while the lower layers remain frozen. This approach preserves generic features while adapting the model to the specific target task.
Fixed Feature Extraction
Method using the lower layers of a pre-trained model as a static extractor without weight modification during training. This technique ensures stability of extracted features while reducing computational costs.
Lower Network Layers
First layers of a deep neural network specialized in detecting elementary and generic patterns. These layers capture universal features transferable between different tasks and application domains.
Visual Descriptors
Quantitative features extracted from images by the lower convolutional layers of a pre-trained model. These descriptors represent fundamental visual attributes like edges, textures, and geometric structures.
Latent Representations
Abstract encodings generated by the hidden layers of a pre-trained neural network that capture essential data information. These representations serve as a foundation for downstream tasks by reducing dimensionality while preserving semantics.
Hierarchical Features
Structure of features organized in increasing levels of abstraction produced by successive layers of a deep network. Lower layers generate reusable low-level features as primitives for various tasks.
Convolutional Feature Maps
Outputs of convolutional layers representing spatial activations corresponding to the presence of specific patterns. These feature maps of lower layers are particularly effective at capturing reusable local structures.
Bottleneck Features
Compressed representations extracted just before the classification layers of a pre-trained network. These bottleneck features capture essential semantic information in a compact format ideal for transfer learning.
Pre-trained Models
Deep learning architectures with weights already optimized on benchmark datasets like ImageNet or COCO. These models provide ready-to-use feature extractors for various vision or language processing tasks.
Transfer Learning
Machine learning paradigm that reuses knowledge acquired on a source task to improve performance on a target task. This approach is particularly effective with feature extraction from lower layers of pre-trained models.
Pattern Extraction
Process of automatic detection of recurring structures in data through filters of lower convolutional layers. Extracted patterns serve as fundamental building blocks for more complex representations.
Abstract Representations
High-level encodings generated by intermediate layers of a network capturing semantic concepts rather than raw pixels. These representations allow better generalization between different related tasks.
Extraction Layers
Set of neural layers specialized in transforming raw data into usable features. In the context of transfer learning, these layers typically come from the lower levels of pre-trained models.
Convolutional Features
Features extracted through convolution operations applied to input data via learned filters. These features from lower layers are particularly effective at capturing translation-invariant local structures.
Pre-trained Embeddings
Dense vector representations generated by models previously trained on large data corpora. These embeddings capture rich semantic relationships and can serve as initial features for specialized tasks.