Learning device and learning method
AssigneeRicoh Company, Ltd.
InventorsTakuya Tanaka, Ryosuke Kasahara
Life of this patent
Estimating the term…
Abstract
A learning device is configured to perform learning of a decision tree by gradient boosting. The learning device includes a plurality of learning units and a plurality of model memories. The plurality of learning units are configured to perform learning of the decision tree using learning data divided to be stored in a plurality of data memories. The plurality of model memories are each configured to store data of the decision tree learned by corresponding one of the plurality of learning units.
Claims10
A learning device configured to perform learning of a decision tree by gradient boosting, the learning device comprising:
- a plurality of learners configured to perform learning of the decision tree using learning data divided into a plurality of pieces to be stored in a plurality of data memories;
- and
- a plurality of model memories each configured to store data of the decision tree learned by a corresponding one of the plurality of learners;
- a discriminator configured to read out each feature amount of a piece of the learning data from the respective data memory, and based on a branch condition for a node of the decision tree, the branch condition being derived based on the feature amount, discriminate a lower node to which the piece of the learning data read out from the respective data memory is to branch from the node;
- wherein the discriminator is configured to update a sample weight by obtaining a sum total of leaf weight of a leaf to which the piece of the learning data is stored and branched in the decision tree;
- wherein the discriminator is configured to update gradient information of the piece of the learning data based on a sum total of leaf weight;
- wherein the discriminator comprises performance calculating circuitry configured to calculate an index value of recognition performance of the learned decision tree based on all sample weights for the plurality of pieces of the learning data learned in parallel;
- and
- the learning device further comprises a plurality of the performance calculating circuitry, wherein each of the plurality of the performance calculating circuitry is configured to calculate the index value based on (i) a label of corresponding one of pieces of the learning data stored in corresponding one of the plurality of data memories, and (ii) the sample weights for the plurality of pieces of the learning data.
A learning method for a learning device configured to perform learning of a decision tree by gradient boosting, the learning method comprising:
- learning the decision tree using learning data divided into a plurality of pieces to be stored in a plurality of data memories by a plurality of learners;
- and
- storing each piece of data of the learned decision tree by corresponding one of a plurality of model memories;
- reading out each feature amount of a piece of the learning data from the respective data memory, and based on a branch condition for a node of the decision tree, the branch condition being derived based on the feature amount, discriminating a lower node to which the piece of the learning data read out from the respective data memory is to branch from the node;
- updating a sample weight by obtaining a sum total of leaf weight of a leaf to which the piece of the learning data is stored and branched in the decision tree;
- updating gradient information by obtaining a sum total of leaf weights of leaves in the decision tree stored in the model memory;
- calculating an index value of recognition performance of the learned decision tree based on all sample weights for the plurality of pieces of the learning data learned in parallel;
- and
- calculating the index value based on (i) a label of corresponding one of pieces of the learning data stored in corresponding one of the plurality of data memories, and (ii) the sample weights for the plurality of pieces of the learning data.
Dependencies are inferred from the claim text.