Active learning framework for machine-assisted tasks

A method for machine learning for a machine-assisted inspection task comprising:

• collecting a set of unlabeled video data observations for object detection;
• performing iterations of an active learning phase followed by a training phase, wherein:
• performing each iteration of the active learning phase includes:
• generating, with a plurality of machine learning components, a plurality of scores for an unlabeled observation of the set of unlabeled video data observations, wherein each machine learning component is configured to generate a different score, of the plurality of scores, for the unlabeled observation, wherein the plurality of scores includes an ensemble disparity score, wherein generating the ensemble disparity score comprises:
• predicting, by each object detector of a plurality of object detectors of the plurality of machine learning components, a bounding box corresponding to a rectangular area in pixel space of a video frame of video data, the bounding box associated with one or more label classes;
• predicting, by each object detector of the plurality of object detectors, a multi-class channel stack, each element in the multi-class channel stack corresponding to a pixel, a label class, and the object detector;
• comparing the elements of the multi-class channel stacks of the plurality of object detectors to compute a disparity value;
• and
• generating the ensemble disparity score based on a weighted sum of the disparity value over the pixel space of the video frame;
• inputting the unlabeled observation to a computation model of a primary machine learning component of the plurality of machine learning components;
• predicting, with the computation model of the primary machine learning component, a label associated with the unlabeled observation, wherein the computation model of the primary machine learning component generates at least one uncertainty score, of the plurality of scores, associated with the predicted label, the plurality of scores further including at least one of:
• a relevance score for the unlabeled observation, a typicality score for the predicted label, or a density score for the unlabeled observation that characterizes a quality of the unlabeled observation or a quality of the predicted label;
• using the plurality of scores to generate a composite score for the unlabeled observation;
• and
• using the composite score to control sampling of the unlabeled observation for a labeling process that generates and stores a ground truth label for the unlabeled observation;
• and
• performing each iteration of the training phase includes updating the plurality of machine learning components using the ground truth labels generated and stored by previous iterations of the active learning phase;
• and
• controlling the machine-assisted task based on labeled observations.