Back to results
US 12,614,382 B1Granted
Neural network based vision systems
AssigneeNVIDIA Corporation
InventorMichael Ranzinger
Life of this patent
Estimating the term…
Abstract
Apparatuses, systems, and techniques to train one or more neural networks using unannotated images. In at least one embodiment, the one or more neural networks are trained based, at least in part, on one or more loss functions calculated using a randomly selected portion pair from two different images and a randomly selected portion pair from the same image.
Claims20
A processor comprising one or more circuits to:
- randomly split each unannotated image from a plurality of images into respective non-overlapping portions of patches;
- randomly select a first pair of two non-overlapping portions respectively from each image of two or more different images of the plurality of images, and a second pair of two non-overlapping portions from a same image;
- encode, by one or more neural networks, patches in the first pair and the second pair into embeddings of patches in an embedding space;
- and
- update one or more parameters of one or more neural networks based, at least in part, on one or more loss functions calculated using first embeddings of patches corresponding to the first randomly selected pair of portions and second embeddings of patches corresponding to the second pair of portions.
2 more elements
A system comprising:
- one or more memories to store one or more neural networks for image tasks and a plurality of processor-executable instructions;
- a communication interface receiving a plurality of images;
- one or more processors executing the processor-executable instructions to:
- randomly split each unannotated image from the plurality of images into respective non-overlapping portions of patches;
- randomly select a first pair of two non-overlapping portions respectively from each image of two or more different images of the plurality of images, and a second pair of two non-overlapping portions from a same image;
- encode, by one or more neural networks, patches in the first pair and the second pair into embeddings of patches in an embedding space;
- and
- update one or more parameters of one or more neural networks based, at least in part, on one or more loss functions calculated using first embeddings of patches corresponding to the first randomly selected pair of portions and second embeddings of patches corresponding to the second pair of portions.
5 more elements
A method comprising:
- randomly splitting each unannotated image from a plurality of images into respective non-overlapping portions of patches;
- randomly selecting a first pair of two non-overlapping portions respectively from each image of two or more different images of the plurality of images, and a second pair of two non-overlapping portions from a same image;
- encoding, by one or more neural networks, patches in the first pair and the second pair into embeddings of patches in an embedding space;
- and
- updating one or more parameters of one or more neural networks based, at least in part, on one or more loss functions calculated using first embeddings of patches corresponding to the first randomly selected pair of portions and second embeddings of patches corresponding to the second pair of portions.
2 more elements
Dependencies are inferred from the claim text.