US-20260127864-A1 - IMAGE-FOR-TRAINING SELECTING APPARATUS, IMAGE-FOR-TRAINING SELECTING METHOD, AND STORAGE MEDIUM FOR DECISION MAKING

US20260127864A1US 20260127864 A1US20260127864 A1US 20260127864A1US-20260127864-A1

Abstract

An image-for-training selecting apparatus for suitably selecting an image-for-training for training a machine learning model includes at least one processor executing: a first training process of training, by contrastive learning using an images-for-training set, a first machine learning model including a first layer group; a second training process of training a second machine learning model including the first layer group and a second layer group and employing the first machine learning model as a pre-trained model; a first calculating process of calculating a first similarity between a parameter of the first layer group after training by the first training process but before training by the second training process and a parameter of the first layer group after training by the second training process; and a first determining process of determining, based on the first similarity, whether the images-for-training set includes an inappropriate image-for-training.

Inventors

Yasuo Omi

Assignees

NEC CORPORATION

Dates

Publication Date: 20260507
Application Date: 20251218

Claims (20)

1 . An image-for-training selecting apparatus comprising at least one processor, the at least one processor executing: a first training operation to train, by contrastive learning, a first machine learning model comprising a feature extraction part which receives input of an image and generates features of the image, the contrastive learning using a set of images-for-training, which is a plurality of images-for-training; a second training operation to train, with use of the set of images-for-training, a second machine learning model (i) comprising the feature extraction part and a classification part which is connected to the feature extraction part and which receives input of the features of the image and classifies the image and (ii) employing the first machine learning model as a pre-trained model; a computing operation to compute a parameter stability metric, the metric representing a similarity between (i) a parameter of the feature extraction part after the first training operation but before the second training operation and (ii) a parameter of the feature extraction part after the second training operation; and an assessing operation to assess, based on the parameter stability metric, whether a potential defect exists in the set of images-for-training.
2 . The image-for-training selecting apparatus according to claim 1 , wherein the at least one processor further executes: a selecting operation to select, as the set of images-for-training, some of a plurality of available images-for-training, and wherein in a case where it is assessed that a potential defect exists, a set of images-for-training which is different from the set of images-for-training having been selected is selected in the selecting operation.
3 . The image-for-training selecting apparatus according to claim 1 , wherein the computing operation comprises computing second parameter stability metrics for respective layers of the feature extraction part.
4 . The image-for-training selecting apparatus according to claim 3 , wherein the computing operation comprises computing, as the parameter stability metric, a value given by dividing a sum of the second parameter stability metrics by the number of the layers in the feature extraction part.
5 . The image-for-training selecting apparatus according to claim 3 , wherein the computing operation comprises computing, as the parameter stability metric, a value given by dividing a weighted sum, which is a sum of the second parameter stability metrics having been given weights, by a sum of values of the weights.
6 . The image-for-training selecting apparatus according to claim 5 , wherein the computing operation comprises giving a heavier weight value to a second parameter stability metric of a layer closer to an output of the first machine learning model.
7 . The image-for-training selecting apparatus according to claim 2 , wherein the at least one processor further executes: a check operation to calculate an imbalance index for attributes of the plurality of images-for-training included in the selected set of images-for-training; and a second assessing operation to assess, based on the imbalance index, whether an attribute imbalance exists.
8 . The image-for-training selecting apparatus according to claim 7 , wherein in a case where it is assessed, in the second assessing operation, that an attribute imbalance exists, the at least one processor selects, in the selecting operation, a set of images-for-training which is different from the set of images-for-training having been selected.
9 . The image-for-training selecting apparatus according to claim 7 , wherein the attributes comprise at least one selected from the group consisting of a facility where an image was captured, a model of an image-capturing apparatus, and a type of a subject included in an image.
10 . The image-for-training selecting apparatus according to claim 2 , wherein the operations are repeatedly performed for a plurality of different sets of images-for-training, and wherein the at least one processor outputs a set of images-for-training corresponding to a highest parameter stability metric among sets for which it is assessed that no potential defect exists.
11 . An image-for-training selecting method comprising: performing a first training operation to train, by contrastive learning, a first machine learning model comprising a feature extraction part which receives input of an image and generates features of the image, the contrastive learning using a set of images-for-training, which is a plurality of images-for-training; performing a second training operation to train, with use of the set of images-for-training, a second machine learning model (i) comprising the feature extraction part and a classification part which is connected to the feature extraction part and which receives input of the features of the image and classifies the image and (ii) employing the first machine learning model as a pre-trained model; computing a parameter stability metric, the metric representing a similarity between (i) a parameter of the feature extraction part after the first training operation but before the second training operation and (ii) a parameter of the feature extraction part after the second training operation; and assessing, based on the parameter stability metric, whether a potential defect exists in the set of images-for-training.
12 . The image-for-training selecting method according to claim 11 , further comprising: selecting, as the set of images-for-training, some of a plurality of available images-for-training; and in a case where it is assessed that a potential defect exists, selecting a set of images-for-training which is different from the set of images-for-training having been selected.
13 . The image-for-training selecting method according to claim 11 , wherein the step of computing the parameter stability metric comprises computing second parameter stability metrics for respective layers of the feature extraction part.
14 . The image-for-training selecting method according to claim 13 , wherein the step of computing the parameter stability metric comprises computing a value given by dividing a weighted sum, which is a sum of the second parameter stability metrics having been given weights, by a sum of values of the weights.
15 . The image-for-training selecting method according to claim 14 , wherein the step of computing the parameter stability metric comprises giving a heavier weight value to a second parameter stability metric of a layer closer to an output of the first machine learning model.
16 . The image-for-training selecting method according to claim 12 , further comprising: calculating an imbalance index for attributes of the plurality of images-for-training included in the selected set of images-for-training; and assessing, based on the imbalance index, whether an attribute imbalance exists.
17 . A non-transitory computer-readable storage medium storing a program that, when executed by a computer, causes the computer to perform a method, the method comprising: performing a first training operation to train, by contrastive learning, a first machine learning model comprising a feature extraction part which receives input of an image and generates features of the image, the contrastive learning using a set of images-for-training, which is a plurality of images-for-training; performing a second training operation to train, with use of the set of images-for-training, a second machine learning model (i) comprising the feature extraction part and a classification part which is connected to the feature extraction part and which receives input of the features of the image and classifies the image and (ii) employing the first machine learning model as a pre-trained model; computing a parameter stability metric, the metric representing a similarity between (i) a parameter of the feature extraction part after the first training operation but before the second training operation and (ii) a parameter of the feature extraction part after the second training operation; and assessing, based on the parameter stability metric, whether a potential defect exists in the set of images-for-training.
18 . The non-transitory computer-readable storage medium according to claim 17 , the method further comprising: selecting, as the set of images-for-training, some of a plurality of available images-for-training; and in a case where it is assessed that a potential defect exists, selecting a set of images-for-training which is different from the set of images-for-training having been selected.
19 . The non-transitory computer-readable storage medium according to claim 17 , wherein the step of computing the parameter stability metric comprises computing a value given by dividing a weighted sum, which is a sum of second parameter stability metrics for respective layers of the feature extraction part having been given weights, by a sum of values of the weights.
20 . The non-transitory computer-readable storage medium according to claim 18 , the method further comprising: calculating an imbalance index for attributes of the plurality of images-for-training included in the selected set of images-for-training; and assessing, based on the imbalance index, whether an attribute imbalance exists.

Description

This application is a Continuation of U.S. application Ser. No. 18/554,752 filed on Oct. 10, 2023, which is a National Stage Entry of PCT/JP2023/001612 filed on Jan. 20, 2023, the contents of all of which are incorporated herein by reference, in their entirety. TECHNICAL FIELD The present invention relates to an image-for-training selecting apparatus, an image-for-training selecting method, and a storage medium for selecting an image-for-training for use in training of a machine learning model. BACKGROUND ART There has been disclosed a technique of selecting an image-for-training for use in training of a machine learning model. Patent Literature 1 discloses a training apparatus including a first training means that executes a first training process of training, by machine learning using training data, a first model that determines a category of given data. Further, the training apparatus disclosed in Patent Literature 1 selects upper-level training data as first training data and lower-level training data as second training data, from among pieces of training data sorted in ascending order of a difference between a determination result given by the first training means and a correct category set by a user. The training apparatus disclosed in Patent Literature 1 further includes a second training means that executes a second training process of learning, by machine learning using the first training data and the second training data, a second learning model that evaluates the training data. CITATION LIST Patent Literature [Patent Literature 1]International Publication No. WO 2019/187594 SUMMARY OF INVENTION Technical Problem However, if the correct category is incorrect, the training apparatus disclosed in Patent Literature 1 cannot appropriately select training data, disadvantageously. The correct category is set by the user, and, in some cases, the user may set a correct category which is incorrect. Further, in a case where setting of a correct category depends on the skill of a person who sets the correct category, e.g., in a case of using pathological cells, the correct category is not always set correctly. Further, in machine learning, it is preferable that training data be balanced and be comprehensive. However, in a case where imbalance is present in training data, the training apparatus disclosed in Patent Literature 1 cannot select inappropriate training data. An example aspect of the present invention was made in consideration of the above problem. An example object of the present invention is to provide a technique for suitably selecting an image-for-training for use in training of a machine learning model. Solution to Problem An image-for-training selecting apparatus in accordance with an example aspect of the present invention includes at least one processor, the at least one processor executing: a first training process of training, by contrastive learning, a first machine learning model including a first layer group which receives input of an image and generates features of the image, the contrastive learning using a set of images-for-training, which is a plurality of images-for-training; a second training process of training, with use of the set of images-for-training, a second machine learning model (i) including the first layer group and a second layer group which is connected to the first layer group and which receives input of the features of an image and classifies the image and (ii) employing the first machine learning model as a pre-trained model; a first calculating process of calculating a first similarity, which is a similarity between (i) a parameter of the first layer group after training by the first training process but before training by the second training process and (ii) a parameter of the first layer group after training by the second training process; and a first determining process of determining, on a basis of the first similarity, whether or not the set of images-for-training includes an inappropriate image-for-training. An image-for-training selecting method in accordance with an example aspect of the present invention includes at least one processor carrying out: training, by contrastive learning, a first machine learning model including a first layer group which receives input of an image and generates features of the image, the contrastive learning using a set of images-for-training, which is a plurality of images-for-training; training, with use of the set of images-for-training, a second machine learning model (i) including the first layer group and a second layer group which is connected to the first layer group and which receives input of the features of the image and classifies the image and (ii) employing the first machine learning model as a pre-trained model; calculating a first similarity, which is a similarity between (i) a parameter of the first layer group after training by the contrastive learning but before training of the second machine learning model