公开(公告)号：US20180039863A1

公开(公告)日：2018-02-08

申请号：US15548396

申请日：2015-03-26

Applicant: Intel Corporation

Inventor： Ke Ding ,  Chun Luo

IPC: G06K9/62 ,  G06N3/063 ,  G06N3/08

CPC classification number: G06K9/6267 ,  G06K9/6232 ,  G06N3/0454 ,  G06N3/063 ,  G06N3/082

Abstract: A classification system is described which may include neural network decomposition logic (“NND”), which may perform classification using a neural network (“NN”). The NND may decompose a classification decision into multiple sub-decision spaces. The NND may perform classification using an NN that has fewer neurons than the NND utilizes for classification and/or which accepts feature vectors of a smaller size than are input into the NND. The NND may maintain multiple contexts for sub-decision spaces, and may switch between these in order to perform classification using the sub-decision spaces. The NND may combine results from the sub-decision spaces to decide a classification. By diving the decision into sub-decision spaces, the NND may provide for classification decisions using NNs that might otherwise be unsuitable for a particular classification decisions. Other embodiments may be described and/or claimed.

公开(公告)号：US10915794B2

公开(公告)日：2021-02-09

申请号：US15548396

申请日：2015-03-26

Applicant: Intel Corporation

Inventor： Ke Ding ,  Chun Luo

IPC: G06K9/62 ,  G06N3/08 ,  G06N3/04 ,  G06N3/063

Abstract: A classification system is described which may include neural network decomposition logic (“NND”), which may perform classification using a neural network (“NN”). The NND may decompose a classification decision into multiple sub-decision spaces. The NND may perform classification using an NN that has fewer neurons than the NND utilizes for classification and/or which accepts feature vectors of a smaller size than are input into the NND. The NND may maintain multiple contexts for sub-decision spaces, and may switch between these in order to perform classification using the sub-decision spaces. The NND may combine results from the sub-decision spaces to decide a classification. By diving the decision into sub-decision spaces, the NND may provide for classification decisions using NNs that might otherwise be unsuitable for a particular classification decisions. Other embodiments may be described and/or claimed.