Glosarium AI
Kamus lengkap Kecerdasan Buatan
Squashing Function
Non-linear activation function specific to capsule networks that compresses the length of activation vectors between 0 and 1 while preserving their spatial direction.
Capsule Activation Vector
Multidimensional vector representation where the length indicates the probability of the entity's existence and the orientation encodes its instantiation parameters.
Transformation Matrix
Weight matrix learned by each lower-level capsule to predict the output of higher-level capsules by transforming the spatial properties of the entity.
Primary Capsule Layer
First layer of capsules that converts convolutional feature maps into multidimensional capsule vectors encoding low-level local properties.
Coupling Coefficients
Scalar coefficients iteratively adjusted that determine the influence of each lower-level capsule on each higher-level capsule in the routing process.
Entity Pose
Set of explicit spatial properties such as position, orientation, scale, and deformation encoded in the capsule vector's orientation.
Hierarchical Representation
Organizational structure where low-level capsules detect simple primitives and high-level capsules combine these detections into complex entities.
Capsule Prediction
Vector obtained by multiplying a lower-level capsule's vector by its transformation matrix to predict the expected output of a higher-level capsule.
Coefficient d'accord
Produit scalaire entre la prédiction d'une capsule inférieure et la sortie actuelle d'une capsule supérieure mesurant leur compatibilité spatiale.
Overlap pooling
Technique alternative au max-pooling qui préserve les informations spatiales en utilisant des capsules pour gérer le chevauchement des caractéristiques détectées.