Słownik AI
Kompletny słownik sztucznej inteligencji
Routing-by-Agreement
Iterative algorithm where low-level capsules agree on the best high-level capsule to transmit their information to, dynamically adjusting coupling coefficients at each iteration.
Dynamic Routing
Iterative optimization process that adjusts connection weights between capsules based on their agreement, replacing traditional static pooling with an adaptive mechanism.
Squashing Function
Non-linear activation function specific to capsules that normalizes vector length between 0 and 1 while preserving its direction, representing the probability of an entity's existence.
Coupling Coefficients
Normalized weights that determine the influence of each low-level capsule on high-level capsules, computed by softmax and iteratively updated according to agreement.
Prediction Vectors
Vectors produced by each low-level capsule after transformation by a weight matrix, predicting the potential output of each parent capsule in the routing process.
Agreement Score
Similarity measure calculated by dot product between the predicted vector and the actual output of the parent capsule, used to adjust coupling coefficients.
Primary Capsules
First layer of capsules transforming low-level features from convolutions into activation vectors, preserving spatial and rotational properties.
Digit Capsules
High-level capsules specialized in final classification, whose length represents the probability of presence of a specific class.
Vector Length Representation
Encoding where the magnitude of the capsule vector indicates the probability of an entity's existence, while its direction encodes its instantiation properties.
Pose Matrices
4x4 affine transformations encoding the position, orientation, and scale of an object, allowing capsules to explicitly represent spatial transformations.
Transformation Matrix
Weight matrix learned by each low-level capsule that transforms its input vector into a prediction vector for each potential parent capsule.
EM Routing
Routing variant using the Expectation-Maximization algorithm to model the distribution of capsule activations, offering better stability and computational efficiency.
Cluster Routing
Routing approach where similar capsules form dynamic clusters, enabling more robust aggregation of partial features into coherent entities.
Capsule Equivariance
Fundamental property where capsule activations vary systematically with input transformations, enabling recognition invariant to pose changes.
Iterative Refinement
Cyclic process of improving coupling coefficients and capsule activations, converging toward an optimal distribution of information between layers.
Routing Iterations
Number of successive steps in the routing algorithm, typically 3 iterations being sufficient to achieve stable convergence of capsule activations.