Techniques for improving data efficiency in image-based machine learning
Inventors
Ozdemir, Oscar • GARDNER, MARC-ANDRÉ • Jahjah, Karl-Alexandre • LAPOINTE, Marc-André • Asselin, Louis-Philippe
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-12682617-B1
Publication Date
2026-07-14
Expiration Date
Abstract
In example embodiments, a dataset labeling software process combines diversity selection and with one or more uncertainty selection techniques in a multi-stage process that produces a labeled training dataset for training an image-based ML model. In an initial stage, diversity selection alone may be used to seed a labeled training dataset. In a subsequent stage, diversity selection may be combined with a first procedure for computing uncertainty (e.g., uncertainty selection based on ensemble learning) to build the labeled training dataset until a first stopping condition is met. Optionally, in a still further stage, diversity selection may be combined with a second, different procedure for computing uncertainty (e.g., uncertainty selection based on Monte Carlo (MC) dropout) to further build the labeled training dataset until a second stopping condition is met.
Core Innovation
A software process executing on one or more computing devices computes a dissimilarity between each image of an initial set of images available for training. The process selects a first sub-set of images from the initial set based on the dissimilarity from each other, labels the first sub-set, and adds the labeled first sub-set to a labeled training dataset. The process then trains one or more image-based machine learning models on the labeled training dataset.
In a subsequent stage, the software process computes an uncertainty of the one or more image-based machine learning models for remaining images of the initial set. The process selects a second sub-set of images from the remaining images based on a combination of the dissimilarity from each other and the uncertainty of the one or more image-based machine learning models. The second sub-set is labeled and added to the labeled training dataset, and the training and uncertainty computation repeat until a stopping condition is met to build the labeled training dataset.
The labeled training dataset is then used to train the image-based machine learning model. The data labeling software process is configured for an initial stage selection based on dissimilarity and for a subsequent stage selection based on a combination of dissimilarity and uncertainty, repeating the computation and selection until a stopping condition is met.
Claims Coverage
The independent claims cover an iterative, multi-stage image dataset labeling and training approach with dissimilarity-based initial subset selection and subsequent uncertainty plus dissimilarity selection, repeated until a stopping condition is met. The claim set includes a method claim, a non-transitory computing medium claim, and a computing device claim.
Initial-stage dissimilarity-based subset labeling for dataset construction
compute a dissimilarity between each image of an initial set of images available for training; select a first sub-set of images from the initial set of images based on the dissimilarity from each other; label the first sub-set of images; and add the labeled first sub-set of images to a labeled training dataset.
Subsequent-stage uncertainty plus dissimilarity subset labeling and iterative repetition until stopping
compute an uncertainty of the one or more image-based ML models for remaining images of the initial set of images; select a second sub-set of images from the remaining images of the initial set based on a combination of the dissimilarity from each other and the uncertainty of the one or more image-based ML models; label the second sub-set of images; add the second sub-set of images to the labeled training dataset; repeating the training of the one or more image-based ML models, and the computing of the uncertainty, and the selecting of the second sub-set until a stopping condition is met to build the labeled training dataset.
Using the labeled training dataset to train the image-based ML model
use the labeled training dataset to train the image-based ML model.
Non-transitory computing instructions for iterative labeled training dataset building
a non-transitory computing device readable medium having instructions stored thereon, the instructions when executed by one or more computing devices operable to: in an initial stage, compute a dissimilarity between each image of an initial set of images; select a first sub-set based on the dissimilarity from each other; label the first sub-set; and add the labeled first sub-set to a labeled training dataset; in a subsequent stage, compute an uncertainty for each image of the initial set; select a second sub-set based on a combination of the dissimilarity and the uncertainty; label the second sub-set; add the sub-set to the labeled training dataset; repeat until a stopping condition is met; and use the labeled training dataset to train an image-based ML model.
Computing device with a data labeling software process for multi-stage training subset selection
a computing device configured to train an image-based ML model, including a memory storing software with a data labeling software process configured to: in an initial stage, compute a dissimilarity between each image of an initial set of images available for training; select a first sub-set based on the dissimilarity from each other; provide the first sub-set for training; in a subsequent stage, compute an uncertainty of the one or more image-based ML models for remaining images; select a second sub-set based on a combination of the dissimilarity and the uncertainty; provide the second sub-set for training; and repeat the computation and selection until a stopping condition is met.
Across the independent claims, the inventive concept is a multi-stage process that first selects a dissimilar set of images for labeling and training, then iteratively selects additional images based on a combination of dissimilarity and model uncertainty until a stopping condition is met, and uses the resulting labeled training dataset to train the image-based ML model.
Stated Advantages
Documented Applications
Training an image-based machine learning model using a labeled training dataset built by selecting and labeling first and second sub-sets of images in an initial stage and subsequent stage until a stopping condition is met.
Computing-device-implemented data labeling software process configured to select first and second sub-sets of images for training image-based ML models and repeating until a stopping condition is met.
Interested in licensing this patent?