Device and method for artificial neural network operation
AssigneeShanghai Cambricon Information Technology Co., Ltd.
InventorsShaoli Liu, Yifan Hao, Yunji Chen, Qi Guo, Tianshi Chen
Life of this patent
Estimating the term…
Abstract
Aspects of data modification for neural networks are described herein. The aspects may include a data modifier configured to receive input data and weight values of a neural network. The data modifier may include an input data configured to modify the received input data and a weight modifier configured to modify the received weight values. The aspects may further include a computing unit configured to calculate one or more groups of output data based on the modified input data and the modifier weight values.
Claims17
An apparatus for modifying data for neural networks, comprising:
- a data modifier circuit configured to receive one or more groups of input data, one or more weight values, and one or more connection values,
- wherein the data modifier circuit includes:
- an input data modifier circuit configured to modify the connection values in accordance with the one or more groups of input data, wherein the input data modifier circuit further includes an input data pruning circuit,
- wherein the input data pruning circuit is configured to:
- identify at least one of the one or more groups of input data, absolute values of which are less than or equal to a first threshold value,
- identify at least one of the one or more connection values that correspond to the identified at least one of the one or more groups of input data, and
- set the identified at least one of the one or more connection values to zero, and
- a weight modifier circuit configured to further modify the connection values in accordance with the one or more weight values to generate modified connection values,
- wherein the input data modifier circuit is configured to modify the one or more groups of input data according to the modified connection values and the weight modifier circuit is configured to modify the one or more weight values according to the modified connection values;
- and
- a computing circuit configured to calculate one or more groups of output data based on the modified input data and the modified weight values.
A method for modifying data for neural networks, comprising:
- receiving, by a data modifier circuit, one or more groups of input data, one or more weight values, and one or more connection values;
- modifying, by an input data modifier circuit of the data modifier circuit, the connection values in accordance with the one or more groups of input data;
- wherein the modifying of the connection values includes:
- identifying, by an input data pruning circuit of the input data modifier circuit, at least one of the one or more groups of input data, absolute values of which are less than or equal to a first threshold value,
- identifying, by the input data pruning circuit of the input data modifier circuit, at least one of the one or more connection values that correspond to the identified at least one of the one or more groups of input data, and
- setting, by the input data pruning circuit of the input data modifier circuit, the identified at least one of the one or more connection values to zero;
- further modifying, by a weight modifier circuit of the data modifier circuit, the connection values in accordance with the one or more weight values to generate modified connection values;
- modifying, by the input data modifier circuit, the one or more groups of input data according to the modified connection values;
- modifying, by the weight modifier circuit, the one or more weight values according to the modified connection values;
- and
- calculating, by a computing circuit, one or more groups of output data based on the modified input data and the modified weight values.
Dependencies are inferred from the claim text.