| CPC G06F 18/214 (2023.01) [G06N 3/08 (2013.01); G06N 5/04 (2013.01); G06N 20/00 (2019.01)] | 19 Claims |
|
1. A computer-implemented method comprising:
receiving, at a provider network, a first plurality of images originated by a computing device of a user, wherein the first plurality of images is associated with a first domain;
receiving, at the provider network, data indicating a first plurality of classification labels corresponding to the first plurality of images but not indicating any object detection labels, wherein classification labels indicate a class to which a respective object belongs, and object detection labels indicate a location of an object within a respective image;
receiving, at the provider network, a request to generate a first plurality of object detection labels corresponding to the first plurality of images;
training, by the provider network, a neural network comprising one or more base network layers, one or more classification layers and one or more object detection layers, the training comprising:
training the one or more classification layers based at least in part on use of the first plurality of images, the first plurality of classification labels, a second plurality of images associated with a second, different domain than the first plurality of images, and a second plurality of classification labels corresponding to the second plurality of images, wherein the first domain comprises a plurality of types of entities that are different from types of entities included in the second domain, and wherein a first entity represented within at least one of the first plurality of images is purposefully assigned a corresponding classification label that is of a same class as the classification label of a second, different type of entity represented within one or more of the second plurality of images,
training the one or more object detection layers based at least in part on use of the second plurality of images and a second plurality of object detection labels corresponding to the second plurality of images, and
backpropagating updates to weights of the one or more base network layers based at least in part on the training of the one or more classification layers and the training of the one or more object detection layers;
running, by the provider network, the trained neural network using at least the first plurality of images to yield the first plurality of object detection labels; and
storing, by the provider network, the first plurality of object detection labels to a second storage location.
|