IP Library › Granted Patent US 11,551,084
Granted Patent B2
US 11,551,084 · App. 16/724,109 · Granted Jan 10, 2023

System and method of robust active learning method using noisy labels and domain adaptation

Inventors: Rajshekhar Das (Pittsburgh, PA); Filipe J. Cabrita Condessa (Pittsburgh, PA); Jeremy Zieg Kolter (Pittsburgh, PA)
Assignee: Robert Bosch GmbH
G06N3/08G06F9/4881G06N3/0454G06N20/00
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 11,551,084
App. No.
16/724,109
Granted
Jan 10, 2023
Kind
B2
Abstract

A system and method is disclosed for labeling an unlabeled dataset, with a labeling budget constraint and noisy oracles (i.e. noisy labels provided by annotator), using a noisy labeled dataset from another domain or application. The system and method combine active learning with noisy labels and active learning with domain adaptation to enhance classification performance.

Claims (31)

1. A method that includes an active learning algorithm, comprising:

receiving a noisily labeled source dataset;

applying a robust deep learning algorithm to identify a source classifier for the noisily labeled source dataset;

determining an adapted classifier using a domain discrimination algorithm that operates on the source classifier; and

assigning a label to one or more samples within an unlabeled target dataset based on the active learning algorithm that operably uses a target classifier and the adapted classifier wherein the active learning algorithm is designed using a maximum entropy-based heuristic algorithm.

2. The method of claim 1 , wherein the domain discrimination algorithm is designed as an adversarial network.

3. The method of claim 1 , further comprising: warm-starting the active learning algorithm using the adapted classifier, wherein the active learning algorithm is operable to provide training data for the target classifier.

4. The method of claim 3 , wherein a scheduler controls a duration for warm-starting the active learning algorithm.

5. The method of claim 1 , wherein the robust deep learning algorithm includes a softmax layer followed by a dense linear layer operable to perform denoising of the noisily labeled source dataset.

6. The method of claim 1 , wherein the active learning algorithm is designed using a batch active learning algorithm.

7. The method of claim 6 , wherein the target classifier is updated per-batch.

8. A method that includes an active learning algorithm, comprising:

receiving a noisily labeled source dataset;

applying a robust deep learning algorithm to identify a source classifier for the noisily labeled source dataset;

determining an adapted classifier using an adversarial domain adaptation machine learning algorithm that operates on the source classifier; and

assigning a label to one or more samples within an unlabeled target dataset based on the active learning algorithm that operably uses a target classifier and the adapted classifier.

9. The method of claim 8 , further comprising: warm-starting the active learning algorithm using the adapted classifier, wherein the active learning algorithm is operable to provide training data for the target classifier.

10. The method of claim 9 , wherein a scheduler controls a duration for warm-starting the active learning algorithm.

11. The method of claim 8 , wherein the robust deep learning algorithm includes a softmax layer followed by a dense linear layer operable to perform denoising of the noisily labeled source dataset.

12. A system operable to employ an active learning algorithm, comprising:

a processor operable to:

receive a noisily labeled source dataset;

apply a robust deep learning algorithm to identify a source classifier for the noisily labeled source dataset;

determine an adapted classifier using an domain discrimination machine learning algorithm that operates on the source classifier; and

assign a label to one or more samples within an unlabeled target dataset based on the active learning algorithm that operably uses a target classifier and the adapted classifier wherein the active learning algorithm is designed using a maximum entropy-based heuristic algorithm.

13. The system of claim 12 , wherein the domain discrimination machine learning algorithm is designed using an adversarial network.

14. The system of claim 12 , wherein the processor is further operable to: warm-start the active learning algorithm using the adapted classifier, wherein the active learning algorithm is operable to provide training data for the target classifier.

15. The system of claim 14 , wherein the system further comprises a scheduler that controls a duration to warm-start the active learning algorithm.

16. The system of claim 12 , wherein the robust deep learning algorithm includes a softmax layer followed by a dense linear layer operable to perform denoising of the noisily labeled source dataset.

17. The system of claim 12 , wherein the active learning algorithm is designed using a batch active learning algorithm.

18. The system of claim 17 , wherein the target classifier is updated per-batch.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Dec 20, 2019
From: DAS, RAJSHEKHAR; CABRITA CONDESSA, FILIPE J.; KOLTER, JEREMY ZIEG
To: ROBERT BOSCH GMBH
Reel/Frame 051398/0743 →
Continuity (1)
Related Publication 20210192335A1 · Jun 24, 2021