MULTITASK LEARNING DEVICE, ESTIMATION DEVICE, LEARNING METHOD, AND LEARNING PROGRAM

Title:

MULTITASK LEARNING DEVICE, ESTIMATION DEVICE, LEARNING METHOD, AND LEARNING PROGRAM

Document Type and Number:

WIPO Patent Application WO/2020/230215

Kind Code:

A1

Abstract:

The present invention makes it possible to learn parameters of an estimator using weights determined by taking classes into account.　On the basis of estimation results of tasks, a loss calculation unit calculates, for each task, a loss representing an error in an estimation result of each class for the task. A loss integration unit calculates, for each task, a degree of difficulty for each class, on the basis of the loss calculated for each class for the task. The loss integration unit calculates, for each class, an integrated loss, which is obtained by integrating losses, on the basis of the degree of difficulty calculated for the class for each task and on the basis of the loss calculated for each class for each task. A parameter optimization unit updates the parameters of the estimator on the basis of data, answer data for the data, and the integrated losses. Each process is repeated. Each degree of difficulty is calculated on the basis of a loss change as determined from the loss resulting from a current process in an iteration and the loss resulting from the immediately preceding process in the iteration.

Inventors:

HOSONO TAKASHI (JP)
SHIMAMURA JUN (JP)
SAGATA ATSUSHI (JP)

Application Number:

PCT/JP2019/018832

Publication Date:

November 19, 2020

Filing Date:

May 10, 2019

Export Citation:

Click for automatic bibliography generation Help

Assignee:

NIPPON TELEGRAPH & TELEPHONE (JP)

International Classes:

G06N20/00

Other References:

BUI, MA I ET AL.: "When Regression Meets Manifold Learning for Object Recognition and Pose Estimation", PROCEEDINGS OF 2018 IEEE INTERNATIONAL CONFERENCE ON ROBOTICS AND AUTOMATION (ICRA), 16 May 2018 (2018-05-16), pages 6140 - 6146, XP055759631, ISSN: 2577-087X
OZAN SENER; VLADLEN KOLTUN: "Multi-Task Learning as Multi- Objective Optimization", ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 31(NIPS 2018, 10 October 2018 (2018-10-10), XP080923703

Attorney, Agent or Firm:

TAIYO, NAKAJIMA & KATO (JP)

Download PDF:

View/Download PDF PDF Help

Previous Patent: DEPTH ESTIMATION DEVICE, DEPTH ESTIMATION MODEL LEARNING DEVICE, DEPTH ESTIMATION METHOD, DEPTH ESTI...

Next Patent: CONTROL DEVICE, SYSTEM, METHOD, AND PROGRAM