公开(公告)号：US20240078712A1

公开(公告)日：2024-03-07

申请号：US18306771

申请日：2023-04-25

Applicant: Google LLC

Inventor： David Charles Minnen ,  Saurabh Singh ,  Johannes Balle ,  Troy Chinen ,  Sung Jin Hwang ,  Nicholas Johnston ,  George Dan Toderici

IPC: G06T9/00 ,  G06F17/18 ,  G06N3/08 ,  G06N20/00 ,  G06T3/40

CPC classification number: G06T9/001 ,  G06F17/18 ,  G06N3/08 ,  G06N20/00 ,  G06T3/40 ,  G06T9/002

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for compressing and decompressing data. In one aspect, a method comprises: processing data using an encoder neural network to generate a latent representation of the data; processing the latent representation of the data using a hyper-encoder neural network to generate a latent representation of an entropy model; generating an entropy encoded representation of the latent representation of the entropy model; generating an entropy encoded representation of the latent representation of the data using the latent representation of the entropy model; and determining a compressed representation of the data from the entropy encoded representations of: (i) the latent representation of the data and (ii) the latent representation of the entropy model used to entropy encode the latent representation of the data.

公开(公告)号：US20230186166A1

公开(公告)日：2023-06-15

申请号：US18165211

申请日：2023-02-06

Applicant: Google LLC

Inventor： Deniz Oktay ,  Saurabh Singh ,  Johannes Balle ,  Abhinav Shrivistava

IPC: G06N20/00 ,  G06N3/08

CPC classification number: G06N20/00 ,  G06N3/08

Abstract: Example aspects of the present disclosure are directed to systems and methods that learn a compressed representation of a machine-learned model (e.g., neural network) via representation of the model parameters within a reparameterization space during training of the model. In particular, the present disclosure describes an end-to-end model weight compression approach that employs a latent-variable data compression method. The model parameters (e.g., weights and biases) are represented in a “latent” or “reparameterization” space, amounting to a reparameterization. In some implementations, this space can be equipped with a learned probability model, which is used first to impose an entropy penalty on the parameter representation during training, and second to compress the representation using arithmetic coding after training. The proposed approach can thus maximize accuracy and model compressibility jointly, in an end-to-end fashion, with the rate-error trade-off specified by a hyperparameter.

公开(公告)号：US11574232B2

公开(公告)日：2023-02-07

申请号：US15931016

申请日：2020-05-13

Applicant: Google LLC

Inventor： Deniz Oktay ,  Saurabh Singh ,  Johannes Balle ,  Abhinav Shrivastava

IPC: G06N20/00 ,  G06N3/08

Abstract: Example aspects of the present disclosure are directed to systems and methods that learn a compressed representation of a machine-learned model (e.g., neural network) via representation of the model parameters within a reparameterization space during training of the model. In particular, the present disclosure describes an end-to-end model weight compression approach that employs a latent-variable data compression method. The model parameters (e.g., weights and biases) are represented in a “latent” or “reparameterization” space, amounting to a reparameterization. In some implementations, this space can be equipped with a learned probability model, which is used first to impose an entropy penalty on the parameter representation during training, and second to compress the representation using arithmetic coding after training. The proposed approach can thus maximize accuracy and model compressibility jointly, in an end-to-end fashion, with the rate-error trade-off specified by a hyperparameter.

公开(公告)号：US11538197B2

公开(公告)日：2022-12-27

申请号：US17021688

申请日：2020-09-15

Applicant: Google LLC

Inventor： David Charles Minnen ,  Saurabh Singh

IPC: G06T9/00 ,  G06F17/18 ,  G06N3/04 ,  G06N3/08

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for channel-wise autoregressive entropy models. In one aspect, a method includes processing data using a first encoder neural network to generate a latent representation of the data. The latent representation of data is processed by a quantizer and a second encoder neural network to generate a quantized latent representation of data and a latent representation of an entropy model. The latent representation of data is further processed into a plurality of slices of quantized latent representations of data wherein the slices are arranged in an ordinal sequence. A hyperprior processing network generates a hyperprior parameters and a compressed representation of the hyperprior parameters. For each slice, a corresponding compressed representation is generated using a corresponding slice processing network wherein a combination of the compressed representations form a compressed representation of the data.

公开(公告)号：US20200027247A1

公开(公告)日：2020-01-23

申请号：US16515586

申请日：2019-07-18

Applicant: Google LLC

Inventor： David Charles Minnen ,  Saurabh Singh ,  Johannes Balle ,  Troy Chinen ,  Sung Jin Hwang ,  Nicholas Johnston ,  George Dan Toderici

IPC: G06T9/00 ,  G06T3/40 ,  G06N3/08 ,  G06N20/00 ,  G06F17/18

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for compressing and decompressing data. In one aspect, a method comprises: processing data using an encoder neural network to generate a latent representation of the data; processing the latent representation of the data using a hyper-encoder neural network to generate a latent representation of an entropy model; generating an entropy encoded representation of the latent representation of the entropy model; generating an entropy encoded representation of the latent representation of the data using the latent representation of the entropy model; and determining a compressed representation of the data from the entropy encoded representations of: (i) the latent representation of the data and (ii) the latent representation of the entropy model used to entropy encode the latent representation of the data.

公开(公告)号：US20190356330A1

公开(公告)日：2019-11-21

申请号：US15985340

申请日：2018-05-21

Applicant: Google LLC

Inventor： David Charles Minnen ,  Michele Covell ,  Saurabh Singh ,  Sung Jin Hwang ,  George Dan Toderici

IPC: H03M7/30 ,  G06F17/30 ,  G06N3/08 ,  G06N7/00 ,  H04N19/13 ,  H04N19/192

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for compressing and decompressing data. In one aspect, an encoder neural network processes data to generate an output including a representation of the data as an ordered collection of code symbols. The ordered collection of code symbols is entropy encoded using one or more code symbol probability distributions. A compressed representation of the data is determined based on the entropy encoded representation of the collection of code symbols and data indicating the code symbol probability distributions used to entropy encode the collection of code symbols. In another aspect, a compressed representation of the data is decoded to determine the collection of code symbols representing the data. A reconstruction of the data is determined by processing the collection of code symbols by a decoder neural network.