Słownik AI
Kompletny słownik sztucznej inteligencji
Capsules and Activation Vectors
Neural units representing entity properties as vectors where length indicates presence probability.
Routing-by-Agreement Algorithm
Iterative mechanism allowing low-level capsules to agglomerate towards high-level capsules via dynamic consensus.
Pose Matrices
Learnable transformation matrices encoding spatial and hierarchical relationships between different capsules.
PrimaryCaps Layer
First capsule layer transforming convolutional feature maps into elementary capsule vectors.
DigitCap Layer
Upper capsule layer performing final classification while preserving hierarchical relationships.
Squashing Function
Non-linear activation function specific to capsules that squashes vectors to between 0 and 1 while preserving their direction.
Reconstruction Loss
Regularization term forcing capsules to encode useful information via a reconstruction decoder.
Dynamic Routing
Adaptive process determining how lower-level capsules contribute to higher-level capsules based on their agreement.
Equivariance
Fundamental property of capsules where input changes produce corresponding changes in activations.
CapsNet Architecture
Complete structure of the capsule neural network including initial convolutional layers and hierarchical capsule layers.
Matrix Capsules with EM Routing
Advanced variant using matrix capsules and EM algorithm routing for better performance.
Applications in Computer Vision
Use of capsule networks for object recognition, pose detection, and medical image analysis.