Complexity-based progressive training for machine vision model

Abstract

Methods and systems for training machine vision models (MVMs) with noisy training datasets are described. A noisy set of images is received, wherein labels for some of the images are noisy and / or incorrect. A progressively-sequenced learning curriculum is designed for the noisy dataset, wherein the images that are easiest to learn machine-vision knowledge from are sequenced near the beginning of the curriculum, and the images that are harder to learn machine-vision knowledge from are sequenced later in the curriculum. The MVM is trained via providing the sequenced curriculum to a supervised learning method, so that the MVM learns from the easiest examples first and the harder training examples later, i. e., the MVM progressively accumulates knowledge from simplest to most complex. In orderto sequence the curriculum, the training images are embedded in a feature space and the complexity of each image is determined via density distributions and clusters in the feature space.

Description

Background technique [0001] The routine use of machine vision models (MVMs) implemented by neural networks, such as deep convolutional neural networks (CNNs), has enabled the deployment of various machine vision tasks, such as image classification, object detection, and semantic segmentation, to a limited domain of specific applications . By employing a large number of convolutional layers, such deep CNN-implemented MVMs determine visual, as well as hidden and latent features within an image, and statistically classify the image (or identify objects depicted in the image) based on the identified features. To learn to recognize these features, and to determine their statistical distribution within images depicting similar objects, the network is trained using a large training dataset. To train MVMs, supervised learning methods are usually employed. Such supervised methods utilize a training dataset comprising pre-labeled (ie annotated) images. The markers indicate the correc...

AI Technical Summary

Problems solved by technology

That said, search-based methods may generate "noisy" or "complex" training / validation datasets where labeling lacks quality, consistency, accuracy, and precision

Using such "noisy" (or "complex") training / validation datasets for CNNs often leads to poor performance on machine vision tasks

Therefore, extending this conventional search-based approach to a larger domain of more generalized machine vision tasks has also proven difficult.

Images