Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
PrimaryCap
First layer of capsules in a capsule network that transforms convolutional feature maps into elementary capsule vectors to capture spatial attributes.
Capsule
Vectorial neuronal unit replacing traditional scalar neurons, simultaneously encoding the probability of existence and the instantiation properties of an entity.
Capsule vector
Multidimensional vectorial representation where the norm indicates the probability of existence and the orientation encodes the parameters of the detected entity.
Convolutional feature map
Two-dimensional feature map produced by convolutional layers serving as input to the PrimaryCap layer for transformation into capsules.
DigitCap layer
Final capsule layer producing 16-length vectors for each class, where the norm represents the classification probability.
Routing coefficients
Learnable weights that determine the influence of each lower capsule on the higher capsules during dynamic routing.
Reconstruction layer
Decoder neural network using capsule activations to reconstruct the input image, reinforcing the learning of representations.
Length of capsule
Euclidean norm of the capsule vector representing the probability that the corresponding entity exists in the input.
Vector neuron
Neuron producing a vector output rather than a scalar, capable of simultaneously representing multiple properties of an entity.
Equivariance
Property of capsules where changes in the input lead to corresponding predictable changes in the output vectors.
Primary convolutional layer
Initial convolutional layer extracting basic features before their transformation into capsules by the PrimaryCap layer.
Instantiation parameters
Set of spatial and appearance properties encoded in the orientation of a capsule vector beyond its simple probability of existence.