AI Glossary
The complete dictionary of Artificial Intelligence
Jigsaw Puzzle Learning
Self-supervised learning technique where the model learns to reconstruct an image from its scrambled fragments, thus developing robust visual representations without annotation.
Patch-based Learning
Learning approach that divides images into patches or local regions to enable the model to learn hierarchical and contextual spatial features.
Spatial Reasoning
Model's ability to understand and manipulate spatial relationships between different parts of an image, essential for solving visual puzzles.
Visual Feature Learning
Automatic extraction of relevant visual features from unlabeled images through reconstruction and spatial arrangement tasks.
Context Prediction
Pretext task where the model predicts the spatial context or relative position of patches based on their intrinsic visual content.
Image Reconstruction
Main objective of puzzle-based learning where the neural network must reconstruct the original image by determining the correct order of fragments.
Patch Shuffling
Random shuffling process of image patches creating the puzzle that the model must solve, serving as a self-supervised learning signal.
Visual Correspondence
Establishment of visual relationships between adjacent or complementary patches enabling their coherent arrangement in image space.
Permutation Invariance
Desirable property of the model where the final representation remains stable despite different initial permutations of patches during training.
Visual Attention
Mechanism allowing the model to focus on visually relevant regions of patches to establish meaningful spatial connections.
Contextual Understanding
Capability developed by the model to understand the global context of the image by correctly assembling local fragments.
Patch Matching
Algorithmic process of identifying compatible patches based on visual and textural criteria for optimal reconstruction.
Spatial Coherence
Principle ensuring that adjacent patches in the final reconstruction exhibit natural and logical transitions at the visual level.
Visual Priors
Innate knowledge about the structure of the visual world that the model exploits to efficiently solve image puzzles.
Unsupervised Feature Extraction
Automatic extraction of discriminative features from raw data without labels, facilitated by puzzle reconstruction tasks.
Visual Embedding
Compact vector representation of the visual features of a patch, used to calculate similarities and spatial compatibilities.