Granted Patent B2

US 12,266,174 · App. 17/862,667 · Granted Apr 1, 2025

Inventors: Biplob Debnath (Princeton, NJ); Srimat Chakradhar (Manalapan, NJ); Oliver Po (San Jose, CA); Asim Kadav (Mountain View, CA); Farley Lai (Santa Clara, CA); Farhan Asif Chowdhury (Albuquerque, NM)

Assignee: NEC Corporation

G06V20/41G06N3/08G06V10/764G06V10/82G06V20/46

Abstract

Claims (20)

1. A computer-implemented method of training a neural network, comprising:

training a feature extractor and a classifier, the feature extractor including a time segmentation network, using a first set of training data that includes one or more base cases, wherein the training data includes video and the time segmentation network extracts features from a plurality of frames from the video and uses segmental consensus to generate a feature vector corresponding to the plurality of frames; and

training the classifier with few-shot adaptation using a second set of training data, smaller than the first set of training data, while keeping parameters of the feature extractor constant.

2. The method of claim 1 , wherein the feature extractor is a video feature extractor and the classifier is an action identification model, and wherein the first set of training data includes base case training examples and the second set of training data includes novel training examples for an action that is not represented in the first set of training data.

3. The method of claim 1 , wherein the segmental consensus is a pooling operation.

4. The method of claim 1 , wherein the feature extractor includes a temporal shift.

5. The method of claim 4 , wherein the training data includes video, and wherein the temporal shift shifts information for a channel of a first frame of the video to a second frame of the video.

6. The method of claim 1 , wherein the classifier performs linear classification.

7. The method of claim 1 , wherein the classifier performs cosine similarity-based classification.

8. A system for training a neural network, comprising:

a hardware processor; and

a memory that stores a computer program, which, when executed by the hardware processor, causes the hardware processor to:

train a feature extractor and a classifier, the feature extractor including a time segmentation network, using a first set of training data that includes one or more base cases, wherein the training data includes video and the time segmentation network extracts features from a plurality of frames from the video and uses segmental consensus to generate a feature vector corresponding to the plurality of frames; and

train the classifier with few-shot adaptation using a second set of training data, smaller than the first set of training data, while keeping parameters of the feature extractor constant.

9. The system of claim 8 , wherein the feature extractor is a video feature extractor and the classifier is an action identification model, and wherein the first set of training data includes base case training examples and the second set of training data includes novel training examples for an action that is not represented in the first set of training data.

10. The system of claim 8 , wherein the segmental consensus is a pooling operation.

11. The system of claim 8 , wherein the feature extractor includes a temporal shift.

12. The system of claim 11 , wherein the training data includes video, and wherein the temporal shift shifts information for a channel of a first frame of the video to a second frame of the video.

13. The system of claim 8 , wherein the classifier performs linear classification.

14. The system of claim 8 , wherein the classifier performs cosine similarity-based classification.

Assignments (2)

Continuity (3)

References Cited (14)