Search

US-12619681-B2 - Systems and methods for domain-aware classification of unlabeled data

US12619681B2US 12619681 B2US12619681 B2US 12619681B2US-12619681-B2

Abstract

Disclosed is a domain-aware semi-supervised machine learning (“DSSL”) system. The system may determine that each data item of a set of data is associated with a particular domain, may identify a model that associates a set of parameters to a particular classification, and may compare parameters of each data item to the set of parameters included in the model. Based on the comparison, the system may determine that a first subset of the data items is associated with the particular classification, and that a second subset of the data items is not associated with the particular classification. The system may identify parameters of the second subset of data items that are different from the set of parameters in the model, and may perform a set of actions that are associated with the particular classification based on the second subset of data items being associated with the particular classification.

Inventors

  • Leonardo Taccari

Assignees

  • VERIZON PATENT AND LICENSING INC.

Dates

Publication Date
20260505
Application Date
20211008

Claims (20)

  1. 1 . A device, comprising: one or more processors configured to: receive a set of image data that includes a plurality of images; identify a machine learning model that associates a set of parameters to a first classification; compare parameters of each image, of the plurality of images, to the set of parameters included in the machine learning model; determine, based on the comparing, that a first subset of images, of the plurality of images, are associated with the first classification; determine, based on the comparing, that a second subset of images, of the plurality of images, are associated with a second classification and are not associated with the first classification; identify one or more parameters of the second subset of images that are different from the set of parameters included in the machine learning model; modify the machine learning model to associate, based on identifying that the one or more parameters of the second subset of images are different from the set of parameters included in the machine learning model, the one or more parameters of the second subset of images with the first classification; identify a set of actions associated with the first classification, wherein the set of actions include one or more vehicle control actions; receive particular image data captured by a camera associated with a particular vehicle; determine that that the particular image data is associated with the one or more parameters of the second subset of images; determine, based on the modified machine learning model and based on determining that the particular image data is associated with the one or more parameters of the second subset of images, that the particular image data is associated with the first classification; and perform the set of actions, associated with the first classification, based on determining that the particular image data is associated with the first classification, wherein performing the set of actions includes outputting instructions to the particular vehicle to perform the one or more vehicle control actions of the set of actions associated with the first classification.
  2. 2 . The device of claim 1 , wherein the one or more processors are further configured to: train the machine learning model based on parameters of a set of labeled data and a plurality of classifications associated with each data item of the set of labeled data, wherein plurality of classifications comprises the first classification.
  3. 3 . The device of claim 1 , wherein the one or more processors are further configured to: compute an error rate of the machine learning model based on the first classification assigned to the set of image data and probabilities with which the machine learning model classifies the set of image data with the first classification; and select one or more parameters of the set of parameters to adjust based on the error rate.
  4. 4 . The device of claim 1 , wherein the one or more processors are further configured to: define a pseudo-label for the plurality of images of the set of image data based on the first subset of images being classified by the machine learning model to the first classification, the second subset of the plurality of images being classified by the machine learning model to second classification, each image of the first subset of images being associated with a same particular domain, and the first subset of images having a greater quantity of images than the second subset of images.
  5. 5 . The device of claim 4 , wherein defining the pseudo-label comprises: replacing a first label that is defined for the second classification and that is output by the machine learning model for the second subset of the first subset of images with a different second label that is defined for the pseudo-label and the first classification.
  6. 6 . The device of claim 1 , wherein the one or more processors are further configured to: train the machine learning model using a first set of images that are labeled with the first classification, wherein the set of data comprises a different second set of images that are not labeled with a classification.
  7. 7 . The device of claim 6 , wherein the one or more processors are further configured to: modify the machine learning model by expanding a parameterized definition of a particular object represented by the first set of images to include the one or more parameters from a subset of the second set of images that correspond to the second subset of images.
  8. 8 . The device of claim 1 , wherein the one or more processors are further configured to: determine whether the machine learning model satisfies convergence criteria, wherein the convergence criteria sets a threshold error rate for the machine learning model or specifies a number of optimization iterations for the machine learning model; and optimize the machine learning model by adjusting additional parameters of the set of parameters in response to the machine learning model, after said modifying, not satisfying the convergence criteria.
  9. 9 . The device of claim 1 , wherein the one or more processors are further configured to: detect a particular object depicted in the particular set of image data captured by the camera associated with the particular vehicle based on determining that the particular image data is associated with the first classification, wherein the set of actions are associated with detection of the particular object.
  10. 10 . A non-transitory computer-readable medium, storing a plurality of processor-executable instructions to: receive a set of image data that includes a plurality of images; identify a machine learning model that associates a set of parameters to a first classification; compare parameters of each image, of the plurality of images, to the set of parameters included in the machine learning model; determine, based on the comparing, that a first subset of images, of the plurality of images, are associated with the first classification; determine, based on the comparing, that a second subset of images, of the plurality of images, are associated with a second classification and are not associated with the first classification; identify one or more parameters of the second subset of images that are different from the set of parameters included in the machine learning model; modify the machine learning model to associate, based on identifying that the one or more parameters of the second subset of images are different from the set of parameters included in the machine learning model, the one or more parameters of the second subset of images with the first classification; identify a set of actions associated with the first classification, wherein the set of actions include one or more vehicle control actions; receive particular image data captured by a camera associated with a particular vehicle; determine that that the particular image data is associated with the one or more parameters of the second subset of images; determine, based on the modified machine learning model and based on determining that the particular image data is associated with the one or more parameters of the second subset of images, that the particular image data is associated with the first classification; and perform the set of actions, associated with the first classification, based on determining that the particular image data is associated with the first classification, wherein performing the set of actions includes outputting instructions to the particular vehicle to perform the one or more vehicle control actions of the set of actions associated with the first classification.
  11. 11 . The non-transitory computer-readable medium of claim 10 , wherein the plurality of processor-executable instructions further include instructions to: train the machine learning model based on parameters of a set of labeled data and a plurality of classifications associated with each data item of the set of labeled data, wherein plurality of classifications comprises the first classification.
  12. 12 . The non-transitory computer-readable medium of claim 10 , wherein the plurality of processor-executable instructions further include processor-executable instructions to: detect a particular object depicted in the particular set of image data captured by the camera associated with the particular vehicle based on determining that the particular image data is associated with the first classification, wherein the set of actions are associated with detection of the particular object.
  13. 13 . The non-transitory computer-readable medium of claim 10 , wherein the plurality of processor-executable instructions further include processor-executable instructions to: compute an error rate of the machine learning model based on the first classification assigned to the set of image data and probabilities with which the machine learning model classifies the set of image data with the first classification; and select one or more parameters of the set of parameters to adjust based on the error rate.
  14. 14 . A method, comprising: receiving a set of image data that includes a plurality of images; identifying a machine learning model that associates a set of parameters to a first classification; comparing parameters of each image, of the plurality of images, to the set of parameters included in the machine learning model; determining, based on the comparing, that a first subset of images, of the plurality of images, are associated with the first classification; determining, based on the comparing, that a second subset of images, of the plurality of images, are associated with a second classification and are not associated with the first classification; identifying one or more parameters of the second subset of images that are different from the set of parameters included in the machine learning model; modifying the machine learning model to associate, based on identifying that the one or more parameters of the second subset of images are different from the set of parameters included in the machine learning model, the one or more parameters of the second subset of images with the first classification; identifying a set of actions associated with the first classification, wherein the set of actions include one or more vehicle control actions; receiving particular image data captured by a camera associated with a particular vehicle; determining that that the particular image data is associated with the one or more parameters of the second subset of images; determining, based on the modified machine learning model and based on determining that the particular image data is associated with the one or more parameters of the second subset of images, that the particular image data is associated with the first classification; and performing the set of actions, associated with the first classification, based on determining that the particular image data is associated with the first classification, wherein performing the set of actions includes outputting instructions to the particular vehicle to perform the one or more vehicle control actions of the set of actions associated with the first classification.
  15. 15 . The method of claim 14 further comprising: computing an error rate of the machine learning model based on the first classification assigned to the set of image data and probabilities with which the machine learning model classifies the set of image data with the first classification; and selecting one or more parameters of the set of parameters to adjust based on the error rate.
  16. 16 . The method of claim 14 further comprising: defining a pseudo-label for the plurality of images of the set of image data based on the first subset of images being classified by the machine learning model to the first classification, the second subset of the plurality of images being classified by the machine learning model to second classification, each image of the first subset of images being associated with a same particular domain, and the first subset of images having a greater quantity of images than the second subset of images.
  17. 17 . The method of claim 16 , wherein defining the pseudo-label comprises: replacing a first label that is defined for the second classification and that is output by the machine learning model for the second subset of the first subset of images with a different second label that is defined for the pseudo-label and the first classification.
  18. 18 . The method of claim 14 further comprising: training the machine learning model using a first set of images that are labeled with the first classification, wherein the set of data comprises a different second set of images that are not labeled with a classification.
  19. 19 . The method of claim 18 further comprising: modifying the machine learning model by expanding a parameterized definition of a particular object represented by the first set of images to include the one or more parameters from a subset of the second set of images that correspond to the second subset of images.
  20. 20 . The method of claim 14 , further comprising: detecting a particular object depicted in the particular set of image data captured by the camera associated with the particular vehicle based on determining that the particular image data is associated with the first classification, wherein the set of actions are associated with detection of the particular object.

Description

BACKGROUND Supervised machine learning may involve training a model based on a labeled set of example data, and using the model to identify and/or predict labels for unlabeled data that is provided as model input. The predictive accuracy (e.g., the error rate) of the supervised machine learning may be highly dependent on the number of examples and the variety in the examples used to train the model. However, collecting such a large and varied labeled data set may be too costly, time-consuming, and/or infeasible for many practical applications. BRIEF DESCRIPTION OF THE DRAWINGS FIGS. 1A and 1B illustrate an example of one or more iterations of a domain-aware semi-supervised machine learning (“DSSL”) procedure in accordance with some embodiments presented herein; FIG. 2 presents a process for performing a supervised machine learning portion of the DSSL in accordance with some embodiments presented herein; FIG. 3 presents a process for performing the unsupervised machine learning portion of the DSSL in accordance with some embodiments presented herein; FIG. 4 illustrates an example for pseudo-label definition and modeling of unlabeled data in accordance with some embodiments presented herein; FIG. 5 presents a process for controlling a system or device based on actions determined from the DSSL classifications in accordance with some embodiments presented herein; FIG. 6 illustrates an example architecture for generating, training, and/or optimizing a classification model based on a tandem supervised and unsupervised learning in accordance with some embodiments presented herein; FIG. 7 illustrates an example architecture for generating, training, and/or optimizing a classification model based on separate modeling of the labeled and unlabeled data sets in accordance with some embodiments presented herein; FIG. 8 illustrates an example environment, in which one or more embodiments may be implemented; FIG. 9 illustrates an example arrangement of a Radio Access Network (“RAN”), in accordance with some embodiments; and FIG. 10 illustrates example components of one or more devices, in accordance with one or more embodiments described herein. DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS The following detailed description refers to the accompanying drawings. The same reference numbers in different drawings may identify the same or similar elements. Embodiments described herein provide systems and methods for domain-aware semi-supervised machine learning (“DSSL”) to train a model based on parameters and labels from a set of labeled data, and to expand a particular classification within the model to incorporate parameters from a first subset of unlabeled data that the model does not classify with the particular classification when a second subset of unlabeled the model classifies to the particular classification is in the same equivalence class or domain as the first subset of unlabeled data. The systems and methods may include using the resulting model for computer vision, object recognition, feature detection, pattern recognition, and/or other applications in which the expanded definition for the particular classification may be used to identify objects, features, and/or data patterns that include forms, coloring, sizes, postures, movements, expressions, perspectives, and/or other variations not found within the set of labeled data. For example, embodiments described herein may be used in autonomous vehicles or semi-autonomous vehicles, in which image and/or video data captured by one or more cameras of such a vehicle may include unlabeled data (e.g., unlabeled image data). As discussed below, such unlabeled data may be analyzed based on one or more domains associated with the unlabeled data to identify objects, vehicles, road signs, pedestrians, etc. in order to control operation of the vehicle. For example, the vehicle may be controlled to avoid objects in the road (e.g., by applying braking inputs, throttle inputs, steering inputs, etc.), alert a driver or other user of the vehicle, etc. that are identified using techniques described herein. In some embodiments, the domain-aware unsupervised learning may use a model that is trained on the first set of labeled data to determine that a majority of unlabeled data from the same equivalence class may be classified with a particular label from the first set of labeled data. Data of a particular “equivalence class” may include unlabeled data that is generated in sequence or within a particular interval of time, by the same instrument, at the same location, and/or with another type of common set of attributes. As such, a set of unlabeled data items within the same equivalence class may be identified as being associated with the same classification or set of classifications, as certain attributes of the set of unlabeled data may be invariant and/or unchanged among the set of unlabeled data. The domain-aware unsupervised learning may apply a pseudo-label to provide the same cla