Learning device, deduction device, data generation device, learning method, and learning program

Abstract

The present invention constructs a learned model that is robust with respect to a difference in attributes regarding acquisition of data and that involves a relatively low introduction cost. A learning device according to an aspect of the present invention executes, for each learning data set: a first training step for training a second encoding unit and a second meta identification unit so that the identifying result of the second meta identification unit matches meta data; a second training step for training respective encoding units and an deduction unit so that the deduction result of the deduction unit matches correct answer data; a third training step for training a first meta identification unit so that the identifying result of the first meta identification unit matches the meta data; and a fourth training step for training a first encoding unit so that the identifying result of the first meta identification unit does not match the meta data. The third and fourth training steps are alternately executed repeatedly.

Description

technical field [0001] The present invention relates to a learning device, an estimation device, a data generating device, a learning method, and a learning program. Background technique [0002] A system for estimating various events related to traffic is being developed using data obtained from various sensors such as cameras installed on roads. For example, Patent Document 1 proposes a method of predicting demand for a predetermined object, such as emergency vehicles, transportation means, and mobile sales, based on area information indicating area conditions. Specifically, machine learning such as a regression model and a neural network is implemented using learning data composed of area information and positive solution data representing needs related to a predetermined object corresponding to the area information. Using the learned model constructed by this machine learning, demand related to a predetermined object is predicted from the area information showing the st...

AI Technical Summary

Problems solved by technology

[0007] Therefore, it is preferable to take some measures so that the results of learning obtained from the training data can be applied to the acquisition of unknown data with different attributes, rather than complicating the structure of the model in order to accept inputs corresponding to all factors.

Images