KI-Glossar
Das vollständige Wörterbuch der Künstlichen Intelligenz
Neural capsule
Neural computing unit that represents entity properties as a multidimensional vector, encoding both the existence and instantaneous parameters of a feature.
Activation vector
Multidimensional vector representation produced by a capsule, where the magnitude indicates the probability of entity presence and the orientation encodes its instantaneous properties.
Pose agreement
Mechanism for evaluating the consistency between a lower-level capsule's prediction and the activations of higher-level capsules, used for dynamic routing.
Agreement routing
Iterative algorithm that adjusts connection weights between capsules based on the agreement between pose predictions and the actual activations of target capsules.
Pose transformation matrix
Learned matrix that transforms the pose vector of a lower-level capsule into a prediction for a higher-level capsule, encoding hierarchical spatial relationships.
Primary capsule layer
First layer of capsules that transforms features extracted by convolutional layers into multidimensional activation vectors representing basic entities.
Digit capsule layer
Final layer of capsules where the length of activation vectors represents the probability of presence of recognized classes or entities in the input.
Capsule vector length
Euclidean norm of a capsule's activation vector, squashed between 0 and 1 to represent the probability that the corresponding entity is present in the input.
Capsule vector orientation
Direction of the activation vector in multidimensional space, encoding the instantaneous properties and pose parameters of the represented entity.
Routing coefficient
Normalized connection weight between capsules, iteratively adjusted during dynamic routing to maximize agreement between predictions and target activations.
Dynamic routing algorithm
Iterative process that optimizes the distribution of routing coefficients between capsules based on pose agreement, replacing traditional max pooling.
Squash function
Non-linear activation function specific to capsules that preserves the vector direction while compressing its length between 0 and 1 to represent a probability.
Prediction vector
Result of multiplying a lower-level capsule activation vector by its transformation matrix, predicting the output of the corresponding higher-level capsule.
Capsule coupling
Process of weighted aggregation of multiple inputs to a higher-level capsule, where weights are dynamically determined by the routing algorithm.
Equivariant invariance
Property of capsules where changes in the input produce corresponding changes in activation vector orientations without affecting their length.
Capsule hierarchy
Multi-level structure where lower-level capsules detect simple parts and higher-level capsules represent complex objects through aggregation.
Consensus aggregation
Mechanism by which a higher capsule computes its activation by aggregating consistent predictions from lower capsules, forming a consensus on the presence of an entity.