Bandwidth compression for neural network systems
AssigneeQUALCOMM Incorporated
InventorsIn Suk Chong, Xianglin Wang, Cheng-Teh Hsieh
Life of this patent
Estimating the term…
Abstract
Techniques and systems are provided for compressing data in a neural network. For example, output data can be obtained from a node of the neural network. Re-arranged output data having a re-arranged scanning pattern can be generated. The re-arranged output data can be generated by re-arranging the output data into the re-arranged scanning pattern. One or more residual values can be determined for the re-arranged output data by applying a prediction mode to the re-arranged output data. The one or more residual values can then be compressed using a coding mode.
Claims29
A method of compressing data in a neural network, the method comprising:
- obtaining activation data from an intermediate node of the neural network, the neural network including at least one input node, at least one intermediate node, and at least one output node, the at least one intermediate node including the intermediate node;
- generating re-arranged activation data having a re-arranged scanning pattern, wherein the re-arranged activation data is generated by re-arranging the activation data from the intermediate node of the neural network into the re-arranged scanning pattern;
- determining one or more residual values for the re-arranged activation data, the one or more residual values being determined by applying a prediction mode to the re-arranged activation data;
- compressing the one or more residual values using a coding mode;
- storing the compressed one or more residual values;
- generating decompressed activation data using the stored one or more compressed residual values;
- and
- providing, as input to one or more additional intermediate nodes of the neural network, the decompressed activation data.
An apparatus for compressing data in a neural network, comprising:
- a memory configured to store the data;
- and
- a processor configured to:
- obtain activation data from an intermediate node of the neural network, the neural network including at least one input node, at least one intermediate node, and at least one output node, the at least one intermediate node including the intermediate node;
- generate re-arranged activation data having a re-arranged scanning pattern, wherein the re-arranged activation data is generated by re-arranging the activation data from the intermediate node of the neural network into the re-arranged scanning pattern;
- determine one or more residual values for the re-arranged activation data, the one or more residual values being determined by applying a prediction mode to the re-arranged activation data;
- compress the one or more residual values using a coding mode;
- store the compressed one or more residual values;
- generate decompressed activation data using the stored one or more compressed residual values;
- and
- provide, as input to one or more additional intermediate nodes of the neural network, the decompressed activation data.
A non-transitory computer-readable medium having stored thereon instructions that, when executed by one or more processors, cause the one or more processors to:
- obtain activation data from an intermediate node of a neural network, the neural network including at least one input node, at least one intermediate node, and at least one node, the at least one intermediate node including the intermediate node;
- generate re-arranged activation data having a re-arranged scanning pattern, wherein the re-arranged activation data is generated by re-arranging the activation data from the intermediate node of the neural network into the re-arranged scanning pattern;
- determine one or more residual values for the re-arranged activation data, the one or more residual values being determined by applying a prediction mode to the re-arranged activation data;
- compress the one or more residual values using a coding mode;
- store the compressed one or more residual values;
- generate decompressed activation data using the stored one or more compressed residual values;
- and
- provide, as input to one or more additional intermediate nodes of the neural network, the decompressed activation data.
Dependencies are inferred from the claim text.