Low displacement rank based deep neural network compression
AssigneeInterDigital VC Holdings, Inc.
InventorsFabien RACAPE, Swayambhoo JAIN, Shahab HAMIDI-RAD, Dimitris PAPADIMITRIOU
Life of this patent
Estimating the term…
Abstract
A method and an apparatus for performing deep neural network compression use an approximation training set along with information, such as in matrices representing weights, biases and non-linearities, to iteratively compress a pre-trained deep neural network by low displacement rank based approximation of the network layer weight matrices. The low displacement rank approximation allows for replacement of an original layer weight matrices of the pre-trained deep neural network as the sum of a small number of structured matrices, allowing compression and low inference complexity.
Claims21
A method, comprising:
- obtaining information representative of weights, biases and non-linearities of a deep neural network, and an approximation training set for the deep neural network;
- generating output matrices of information for a layer of the deep neural network using said information and the approximation training set;
- and,
- updating said output matrices for additional layers by iteratively using output matrices of previous layers of the deep neural network and said approximation training set until a minimization criterion has been met for that layer, and until a last layer is reached to generate coefficients representative of said output matrices;
- and,
- performing quantization and entropy coding on said coefficients.
An apparatus, comprising:
- a processor, configured to perform:
- obtaining matrices representative of weight, biases and non-linearities of a deep neural network;
- generating an output for a layer of the deep neural network using said matrices, an approximation set from training data, and an input to the layer;
- and;
- updating said output matrices for additional layers by iteratively using output matrices of previous layers of the deep neural network and said approximation training set until a minimization criterion has been met for that layer, and until a last layer is reached to generate coefficients representative of said output matrices;
- and,
- performing quantization and entropy coding on said coefficients.
A method, comprising:
- performing entropy decoding on data from a bitstream;
- performing inverse quantization on said entropy decoded data;
- and,
- iteratively obtaining dequantized matrices representative of a deep neural network and building weight matrices until a last layer is reached.
An apparatus, comprising:
- a processor, configured to perform:
- performing entropy decoding on data from a bitstream;
- performing inverse quantization on said entropy decoded data;
- and,
- iteratively obtaining dequantized matrices representative of a deep neural network and building weight matrices until a last layer is reached.
(canceled)
Dependencies are inferred from the claim text.