Machine learning system and machine learning method
A machine learning system and a machine learning method capable of selecting a pretrained model to be used in transfer learning in a short time without actually executing the transfer learning includes a pretrained model acquisition unit which acquires a pretrained model from a pretrained model storage unit storing a plurality of pretrained models obtained by learning a transfer source task under respective conditions; a transfer learning dataset storage unit configured to store dataset related to a transfer target task; a pretrained model adaptability evaluation unit configured to evaluate adaptability of each pretrained model acquired by the pretrained model acquisition unit to the dataset related to the transfer target task; and a transfer learning unit configured to execute, based on an evaluation result of the pretrained model adaptability evaluation unit, transfer learning using a selected pretrained model and the dataset, and outputs a learning result as a trained model.
1 . A machine learning system comprising:
one or more processors; and
a non-transitory computer-readable memory storing instructions that, when executed by the one or more processors, cause the one or more processors to:
access, from a pretrained model repository, a plurality of pretrained models obtained by learning a transfer source task under respective conditions;
access a transfer learning dataset related to a transfer target task, the transfer learning dataset comprising input data and teaching information;
for each pretrained model of the plurality of pretrained models, evaluate an adaptability score of the pretrained model to the transfer learning dataset by:
forward propagating at least a portion of the input data through the pretrained model to obtain an image segmentation result that segments the input data into a plurality of areas such that a same area has a same meaning;
for each segmented area, obtaining a respective feature representation produced by the pretrained model for the segmented area;
determine a correspondence between segmented areas in the image segmentation result and classes indicated by the teaching information based at least in part on an overlap degree between (A) pixels of the segmented areas and (B) pixels associated with the teaching information; and
computing the adaptability score as a function representing a relation between the image segmentation result and the teaching information using at least one of: a probability density, an overlap degree of the segmented areas, and a similarity of the respective feature representations for the segmented areas;
select, based on the adaptability scores, a selected pretrained model from the plurality of pretrained models; and
execute transfer learning using the selected pretrained model and the transfer learning dataset to output a trained model.
2 . The machine learning system according to claim 1 , wherein each pretrained model comprises a neural network including:
a feature extraction subnetwork configured to extract a feature from the input data;
an image segmentation subnetwork configured to generate the image segmentation result; and
a feature prediction subnetwork configured to output a different feature representation for each segmented area.
3 . The machine learning system according to claim 2 , wherein
the adaptability score is computed between the image segmentation result generated for the input data of the transfer learning dataset and the teaching information of the transfer learning dataset.
4 . The machine learning system according to claim 3 , wherein the pretrained model is obtained by executing self-supervised learning on the transfer source task.
5 . The machine learning system according to claim 3 , further comprising:
a display device, wherein the instructions further cause the one or more processors to control the display device to display the image segmentation result and the teaching information together.
6 . The machine learning system according to claim 3 , wherein:
the adaptability score is computed by the function representing the relation between the image segmentation result and the teaching information; and
the function uses at least one of the probability density, the overlap degree of the segmented areas, and the similarity of the respective feature representations for each segmented area.
7 . The machine learning system according to claim 1 , wherein the instructions further:
cause the one or more processors to select the selected pretrained model from the plurality of pretrained models stored in the pretrained model repository.
8 . The machine learning system according to claim 7 , further comprising-a display device, wherein the instructions further cause the one or more processors to control the display device to display the adaptability score for each pretrained model.
9 . The machine learning system according to claim 1 , wherein
evaluating the adaptability score comprises forward propagating the input data of the transfer learning dataset through the pretrained model a number of times less than that at a time of executing the transfer learning.
10 . The machine learning system according to claim 1 , wherein:
the transfer learning dataset comprises an image obtained in an inspection process as the input data; and
the pretrained model repository stores at least one pretrained model obtained by learning a task related to the inspection process.
11 . The machine learning system according to claim 1 , wherein determining the correspondence comprises performing an assignment between segmented areas and classes indicated by the teaching information by selecting, for each class indicated by the teaching information, a segmented area that maximizes an overlap degree with pixels associated with the class, and wherein the overlap degree comprises an intersection over union (IoU).
12 . The machine learning system according to claim 1 , wherein computing the adaptability score comprises, for each segmented area, computing a representative feature as an average of feature values output for pixels belonging to the segmented area, and computing the similarity of the respective feature representations based on distances between the representative features.
13 . The machine learning system according to claim 1 , wherein evaluating the adaptability score comprises forward propagating the input data through each pretrained model only one time per input image or at most ten times per input image.
14 . The machine learning system according to claim 1 , wherein evaluating the adaptability score comprises selecting the portion of the input data by random sampling from the transfer learning dataset.
15 . A machine learning method comprising:
accessing a plurality of pretrained models obtained by learning a transfer source task under respective conditions;
accessing a transfer learning dataset related to a transfer target task, the transfer learning dataset comprising input data and teaching information;
for each pretrained model of the plurality of pretrained models, evaluating an adaptability score of the pretrained model to the transfer learning dataset by forward propagating at least a portion of the input data through the pretrained model to obtain an image segmentation result that segments the input data into a plurality of areas such that a same area has a same meaning;
for each segmented area, obtaining a respective feature representation produced by the pretrained model for the segmented area;
determining a correspondence between segmented areas in the image segmentation result and classes indicated by the teaching information based at least in part on an overlap degree between (A) pixels of the segmented areas and (B) pixels associated with the teaching information; and
computing the adaptability score as a function representing a relation between the image segmentation result and the teaching information using at least one of a probability density, an overlap degree of the segmented areas, and a similarity of the respective feature representations for the segmented areas;
selecting, based on the adaptability scores, a selected pretrained model from the plurality of pretrained models; and
executing, transfer learning using the selected pretrained model and the transfer learning dataset to output a trained model.