Diagonal BiLSTM

E743715

Diagonal BiLSTM is a recurrent neural network architecture used in PixelRNN models to efficiently capture two-dimensional spatial dependencies in images by processing pixels along diagonals with bidirectional LSTMs.

Try in SPARQL Jump to: Statements Referenced by

Statements (45)

Predicate Object
instanceOf bidirectional LSTM variant
component of PixelRNN
assumes fixed raster-scan ordering of pixels
belongsTo autoregressive generative models
deep learning architectures for images
constrains receptive field to previously generated pixels
constrainsDependencies to pixels in previous rows and columns
contrastsWith convolution-only autoregressive models like PixelCNN
ensures no access to future pixels in generation order
hasAdvantage better utilization of 2D structure than simple row-wise RNNs
more parallel computation than fully sequential pixel RNNs
hasComponent bidirectional passes along diagonals
diagonal recurrent connections
hasDirection backward direction along diagonal
forward direction along diagonal
hasProperty bidirectional processing along diagonals
captures long-range spatial dependencies
causal with respect to raster-scan ordering of pixels
parallelizable along image diagonals
uses LSTM gating mechanisms
implementedWith LSTM cells
inputType 2D image grid
inspiredBy sequence modeling with LSTMs
introducedBy Aaron van den Oord NERFINISHED
Koray Kavukcuoglu NERFINISHED
Nal Kalchbrenner NERFINISHED
introducedIn Pixel Recurrent Neural Networks NERFINISHED
operatesOn image pixels
operationalDomain computer vision
generative modeling
optimizationMethod stochastic gradient descent variants
outputType conditional pixel distributions
processes pixels along diagonals
publishedIn ICML 2016
relatedTo PixelCNN NERFINISHED
PixelRNN NERFINISHED
Row LSTM NERFINISHED
autoregressive density estimation for images
trainingMethod maximum likelihood estimation
usedFor autoregressive image modeling
modeling two-dimensional spatial dependencies in images
usedIn PixelRNN NERFINISHED
usedInTask image completion
image density modeling
image generation

Referenced by (1)

Full triples — surface form annotated when it differs from this entity's canonical label.

PixelRNN architectureVariant Diagonal BiLSTM