Edge-based adaptive machine learning for object recognition

Patent metadata

Field	Value
Publication number	US-11205100-B2
Application number	US-201715801773-A
Country	US
Kind code	B2
Filing date	Nov 2, 2017
Priority date	Oct 24, 2016
Publication date	Dec 21, 2021
Grant date	Dec 21, 2021

Abstract

Examples of techniques for adaptive model training are provided. According to one or more embodiments of the present invention, a computer-implemented method for adaptive model training includes generating, by a processing system, a training instance based at least in part on a plurality of images that match a contextual specification of a target visual domain. The method further includes extracting, by the processing system, objects from one of the plurality of images. The method further includes for each extracted object, generating, by the processing system, a plurality of machine learning model features and label recommendations for a user.

First claim

What is claimed is:

1. A computer-implemented method for adaptive model training, the method comprising:
receiving, by a processing system, an adaption task for object recognition at a target visual domain;
determining, by the processing system, whether an existing adaptive model stored in a model repository is trained to recognize objects of the target visual domain, wherein the determining is based on whether a location of the target visual domain overlaps a location of a visual domain of the existing adaptive model;
generating, by the processing system, a training instance based at least in part on visual objects found in a plurality of clustered images stored on a first worker user's electronic device, wherein the clustered images match a contextual specification of the target visual domain, wherein the contextual specification is received from a master user's electronic device via an adaptation task, wherein generating the training instance is based on determining that the the existing adaptive model is not trained to recognize objects of the target visual domain, and wherein generating the training instance comprises:
generating, by the processing system, a plurality of candidate labels, wherein the labels are associated with the target visual domain;
extracting, by the processing system, objects from one of the plurality of images stored on the first worker user's electronic device;
receiving, by the processor system, a recommended label selected from the plurality of candidate labels for each respective extracted object;
extracting, by the processing system, a respective plurality of machine learning model features describing each respective extractive object, from a hidden layer of a machine learning model stored on the first worker user's electronic device, and simultaneous to a selection of the recommended label;
transmitting, by the processor, the respective plurality of extracted machine learning model features and recommended labels to the master user's electronic device to train an adaptive model.

2. The computer-implemented method of claim 1 further comprising, based at least in part on presenting the machine learning model features and label recommendations to a first worker user, receiving, by the processing system, the selected recommended label.

3. The computer-implemented method of claim 2 , wherein the machine learning model features and the selected recommended label form the training instance.

4. The computer-implemented method of claim 3 further comprising training an adapted model corresponding to the target visual domain based at least in part on the machine learning features extracted from the images of first worker user's electronic device, and based at least in part on the selected recommended label.

5. The computer-implemented method of claim 1 , wherein the training instance is stored in an adaptation database.

6. The computer-implemented method of claim 5 , wherein the adapted model is stored in the adaptation database.

7. The computer-implemented method of claim 1 , wherein:
the contextual specification comprises a set of attributes; and
the set of attributes includes one or more of location information, time information, and weather information.

8. The computer-implemented method of claim 1 , wherein generating the plurality of machine learning model features and label recommendations further comprises:
inputting object images into a generic machine learning model;
outputting, for each object image of the object images, a respective output vector represented by the probabilities of the object image belonging to different recognizable classes; and
obtaining a final recognition result via a late fusion technique.

9. The computer-implemented method of claim 1 further comprising:
receiving an adaptation task describing the target visual domain by the set of context attributes; and
determining whether to create the adapted model based at least in part on the target visual domain.

10. The computer-implemented method of claim 1 further comprising:
analyzing an adaptation database to determine whether an adapted model exists for the target visual domain; and
determining whether to create the adapted model based at least in part on the determination of whether an adapted model exists for the target visual domain.

11. The computer-implemented method of claim 10 , comprising requesting the first worker user electronic device assist in creating the adapted model based at least a part on a determination to create the adapted model.

12. The computer-implemented method of claim 1 , wherein the adapted model is a support vector machine (SVM).

Assignees

• IBM

Inventors

• Desai Nirmit V
• Li Dawei
• Salonidis Theodoros

Classifications

• G06V30/19167

  Active pattern learning · CPC title

• G06V30/19173

  Classification techniques · CPC title

• G06V10/82

  using neural networks · CPC title

• G06F18/2178

  based on feedback of a supervisor · CPC title

• G06N20/00 Primary

  Machine learning · CPC title