METHOD OF TRAINING NEURAL NETWORK, AND RECOGNITION METHOD AND APPARATUS USING NEURAL NETWORK

• US 20180174001A1
• Filed: 08/31/2017
• Published: 06/21/2018
• Est. Priority Date: 12/15/2016
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: KANG, Hyoa
• Granted Patent: US 10,902,292 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/217 Validation; Performance evaluation; Active pattern learning techniques
• CPC Class Codes: G06F 18/232 Non-hierarchical techniques
• CPC Class Codes: G06F 18/2414 Smoothing the distance, e.g. radial basis function networks [RBFN]
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 30/19147 Obtaining sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 30/1916 Validation; Performance evaluation
• CPC Class Codes: G06V 30/19173 Classification techniques
• CPC Class Codes: G06V 30/194 References adjustable by an adaptive method, e.g. learning

Abstract

A training method of a neural network, and a recognition method and apparatus using the neural network are disclosed. The recognition method using the neural network includes obtaining a feature vector generated from a hidden layer of the neural network, in response to data being entered to an input layer of the neural network, and determining a reliability of a recognition result for the data using the feature vector and clusters.

Claims

1. A recognition method using a neural network, comprising:

obtaining a feature vector generated from a hidden layer of the neural network, in response to data being entered to an input layer of the neural network; and

determining a reliability of a recognition result for the data using the feature vector and clusters.

2. The recognition method of claim 1, wherein the determining of the reliability comprises:

determining the reliability of the recognition result by calculating a distance between the feature vector and each of the clusters.

3. The recognition method of claim 2, wherein the determining of the reliability comprises:

determining the reliability of the recognition result to be higher, in response to the distance between the feature vector and the clusters being shorter.

4. The recognition method of claim 1, wherein the determining of the reliability comprises:

determining the reliability of the recognition result by calculating a similarity between the feature vector and each of the clusters.

5. The recognition method of claim 1, wherein each of the clusters are determined based on feature vectors of pieces of training data to be used to train the neural network.

6. The recognition method of claim 5, wherein the feature vectors are obtained from the hidden layer of the neural network.

7. The recognition method of claim 5, wherein the each of the clusters comprises information related to any one or any combination of a centroid of the each of the clusters, a boundary information the each of the, a number of pieces of data in the each of the clusters, a distribution for the each of the clusters, an outlier that abnormally deviates from a feature vector distribution of the each of the clusters, and statistical values of feature vectors included in the each of the clusters.

8. A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1.

9. A training method of a neural network, comprising:

obtaining a feature vector generated from a hidden layer of the neural network, in response to training data being received at an input layer of the neural network;

identifying a cluster corresponding to the training data from a plurality of clusters based on the feature vector; and

training the neural network using the identified cluster, in response to an accuracy of recognition for the training data being less than a threshold.

10. The training method of claim 9, wherein the training of the neural network comprises:

training the neural network using training data in the identified cluster.

11. The training method of claim 9, wherein the training of the neural network comprises:

training the neural network by setting a weight of a loss of the training data in the identified cluster to be higher than a weight of a loss of other training data,

wherein the other training data comprises the training data excluding the training data in the identified cluster.

12. The training method of claim 9, wherein the plurality of clusters are determined based on feature vectors of pieces of training data in a training set of the neural network.

13. The training method of claim 12, wherein the plurality of clusters are updated based on a feature vector obtained in the training process, in response to the training of the neural network being performed.

14. The training method of claim 9, wherein the accuracy of the recognition comprises a loss determined based on an estimated value output from the neural network in response to the training data and a correct value corresponding to the training data.

15. The training method of claim 9, wherein the training of the neural network comprises:

training the neural network using a cluster corresponding to validation data having a lowest recognition accuracy among pieces of validation data in a validation set of the neural network.

16. The training method of claim 15, wherein the validation data is different from the training data.

17. A recognition apparatus comprising:

a processor configured to

recognize data using a neural network,

obtain a feature vector generated from a hidden layer of the neural network in response to receiving the data at an input layer of the neural network, and

determine a reliability of a recognition result for the input data using the feature vector and clusters.

18. The recognition apparatus of claim 17, wherein the processor is further configured to determine the reliability of the recognition result by calculating a distance between the feature vector and each of the clusters.

19. The recognition apparatus of claim 17, wherein the processor is further configured to determine the reliability of the recognition result by calculating a similarity between the feature vector and each of the clusters.

20. The recognition apparatus of claim 17, wherein the clusters are determined based on feature vectors of pieces of training data to be used to train the neural network.