Back to results
US 12,614,380 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 patch pair on a same image and a spatial relationship between two patches within the randomly selected patch pair on the same image.
Claims20
A processor comprising one or more circuits to:
- randomly split an unannotated image into non-overlapping portions of patches;
- randomly select a pair of patches from a first non-overlapping portion of the non-overlapping portions of patches and an anchor patch from a second non-overlapping portion of the non-overlapping portions of patches;
- determine a spatial relationship that represents a first distance between one of the two patches to the anchor patch and a second distance between another one of the two patches to the anchor patch;
- and
- update one or more parameters of the one or more neural networks based, at least in part, on one or more loss functions calculated using embeddings corresponding to the randomly selected pair of patches and the anchor patch and the spatial relationship.
2 more elements
A system comprising:
- one or more processors to:
- randomly split an unannotated image into non-overlapping portions of patches;
- randomly select a pair of patches from a first non-overlapping portion of the non-overlapping portions of patches and an anchor patch from a second non-overlapping portion of the non-overlapping portions of patches;
- determine a spatial relationship that represents a first distance between one of the two patches to the anchor patch and a second distance between another one of the two patches to the anchor patch;
- and
- update one or more parameters of the one or more neural networks based, at least in part, on one or more loss functions calculated using embeddings corresponding to the randomly selected pair of patches and the anchor patch and the spatial relationship;
- and
- one or more memories to store the one or more neural networks.
5 more elements
A method, comprising:
- randomly splitting an unannotated image into non-overlapping portions of patches;
- randomly selecting a pair of patches from a first non-overlapping portion of the non-overlapping portions of patches and an anchor patch from a second non-overlapping portion of the non-overlapping portions of patches;
- determining a spatial relationship that represents a first distance between one of the two patches to the anchor patch and a second distance between another one of the two patches to the anchor patch;
- and
- updating one or more parameters of the one or more neural networks based, at least in part, on one or more loss functions calculated using embeddings corresponding to the randomly selected pair of patches and the anchor patch and the spatial relationship.
2 more elements
Dependencies are inferred from the claim text.