IP Library Granted Patent US 12670415
Granted Patent B2
US 12670415 · App. 17/008,005 · Granted Jun 30, 2026

Visual question answering using model trained on unlabeled videos

Inventors: Elad Amrani (Zikhron Yaakov, IL); Rami Ben-Ari (Kiryat-Ono, IL); Daniel Nechemia Rotman (Haifa, IL); Udi Barzelay (Haifa, IL)
Assignee: International Business Machines Corporation
G06N5/04G06F18/217G06N3/08G06V20/46G10L15/26
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 12670415
App. No.
17/008,005
Granted
Jun 30, 2026
Kind
B2
Abstract

An example system includes a processor to learn a shared embedding space on unlabeled videos using speech visual correspondence. The processor can learn a number of additional embeddings including a question plus video embedding and an answer embedding using the shared embedding space to generate a trained visual question answering model. The processor can execute a visual question answering based on the trained visual question answering model.

Claims (26)

1 . A system, comprising a processor configured to:

pre-train a machine learning model to learn a shared embedding space on unlabeled videos using a speech visual correspondence, wherein an ImageNet pre-trained Resnet is used to extract 2D features from the unlabeled videos, and wherein a Kinetics pre-trained ResNeXt is used to extract 3D features from the unlabeled videos, and wherein the shared embedding space comprises a network layer with both speech and vision features in a single feature vector, and wherein training the shared embedding space with vision embeddings embeds similar speech and images within a threshold proximity;

train the machine learning model to learn a plurality of additional embeddings comprising a question plus video embedding and an answer embedding using the shared embedding space to generate a trained visual question answering model, wherein the plurality of additional embeddings comprise word representations learned using a pre-trained word2vec embedding model; and

executing the trained visual question answering model, wherein the executing comprises a nearest neighbor search over a set of predetermined answers in a shared question plus video and answer space.

2 . The system of claim 1 , wherein the shared embedding space comprises coordinated text-video representations.

3 . The system of claim 1 , wherein the processor is configured to train the visual question answering model using a max margin ranking loss function to embed an answer close to a corresponding question plus video in the shared embedding space.

4 . The system of claim 1 , wherein the question plus video embedding and the answer embedding each comprise a single feature vector.

5 . The system of claim 1 , wherein the visual question answering model comprises a neural network model.

6 . The system of claim 1 , wherein the additional embeddings comprise embedding layers of a neural network.

7 . A computer-implemented method, comprising:

pre-training a machine learning model to learn, via a processor, a shared embedding space on unlabeled videos using a speech visual correspondence, wherein an ImageNet pre-trained Resnet is used to extract 2D features from the unlabeled videos, and wherein a Kinetics pre-trained ResNeXt is used to extract 3D features from the unlabeled videos, and wherein the shared embedding space comprises a network layer with both speech and vision features in a single feature vector, and wherein training the shared embedding space with vision embeddings embeds similar speech and images within a threshold proximity;

train the machine learning model to learn, via the processor, a plurality of additional embeddings comprising a question plus video embedding and an answer embedding using the shared embedding space to generate a trained visual question answering model, wherein the plurality of additional embeddings comprise word representations learned using a pre-trained word2vec embedding model; and

executing the trained visual question answering model, wherein the executing comprises a nearest neighbor search over a set of predetermined answers in a shared question plus video and answer space.

8 . The computer-implemented method of claim 7 , wherein learning the question plus video embedding comprises training the visual question answering model on labeled data.

9 . The computer-implemented method of claim 7 , wherein training the visual question answering model comprises using a max margin ranking loss function to embed an answer with a reduced distance to a corresponding question plus video in the shared embedding space.

10 . The computer-implemented method of claim 7 , wherein learning the question plus video embedding comprises embedding question plus video embeddings with a reduced distance to corresponding answer embeddings.

11 . The computer-implemented method of claim 7 , wherein executing the visual question answering comprises executing an action recognition or object recognition using the trained visual question answering model.

12 . The computer-implemented method of claim 7 , wherein executing the visual question answering comprises executing an image captioning or video captioning using the trained visual question answering model.

13 . A computer program product for training visual question answering models, the computer program product comprising a computer-readable storage medium having program code embodied therewith, the program code executable by a processor to cause the processor to:

pre-train a machine learning model to learn a shared embedding space on unlabeled videos using a speech visual correspondence, wherein an ImageNet pre-trained Resnet is used to extract 2D features from the unlabeled videos, and wherein a Kinetics pre-trained ResNeXt is used to extract 3D features from the unlabeled videos, and wherein the shared embedding space comprises a network layer with both speech and vision features in a single feature vector, and wherein training the shared embedding space with vision embeddings embeds similar speech and images within a threshold proximity;

train the machine learning model to learn a plurality of additional embeddings comprising a question plus video embedding and an answer embedding using the shared embedding space to generate a trained visual question answering model, wherein the plurality of additional embeddings comprise word representations learned using a pre-trained word2vec embedding model; and

executing the trained visual question answering model, wherein the executing comprises a nearest neighbor search over a set of predetermined answers in a shared question plus video and answer space.

14 . The computer program product of claim 13 , further comprising program code executable by the processor to train the visual question answering model on labeled data.

15 . The computer program product of claim 13 , further comprising program code executable by the processor to use a max margin ranking loss function to embed an answer with a reduced distance to a corresponding question plus video in the shared embedding space.

16 . The computer program product of claim 13 , further comprising program code executable by the processor to embed question plus video embeddings with a reduced distance to corresponding answer embeddings.

17 . The computer program product of claim 13 , further comprising program code executable by the processor to execute a computer vision task based on the trained visual question answering model.