US 11,860,977 B1
Hierarchical graph neural networks for visual clustering
Yifan Xing, Bellevue, WA (US); Tianjun Xiao, Nanjing (CN); Tong He, Shanghai (CN); Yongxin Wang, Seattle, WA (US); Yuanjun Xiong, Seattle, WA (US); Wei Xia, Seattle, WA (US); David Paul Wipf, Jing'An (CN); Zheng Zhang, Shanghai (CN); and Stefano Soatto, Pasadena, CA (US)
Assigned to Amazon Technologies, Inc., Seattle, WA (US)
Filed by Amazon Technologies, Inc., Seattle, WA (US)
Filed on May 4, 2021, as Appl. No. 17/307,701.
Int. Cl. G06F 18/2323 (2023.01); G06N 20/00 (2019.01); G06F 18/2415 (2023.01); G06F 18/23213 (2023.01); G06F 18/2413 (2023.01)
CPC G06F 18/2323 (2023.01) [G06F 18/23213 (2023.01); G06F 18/2415 (2023.01); G06F 18/24147 (2023.01); G06N 20/00 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method comprising:
generating a k-nearest neighbors graph comprising a first plurality of nodes connected by a first plurality of edges for a set of visual embeddings from a dataset of images where k is a positive integer;
generating, by a machine learning model for an input of the k-nearest neighbors graph, a linkage probability for each of the first plurality of edges that indicates a probability that an edge is linking nodes that have a same class;
generating, by the machine learning model, a density probability for each of the first plurality of nodes that indicates a probability that a node is a center of a class of its k-nearest neighbors;
merging the first plurality of nodes into a second plurality of nodes, having fewer nodes than the first plurality of nodes, based at least in part on the linkage probabilities and the density probabilities; and
generating a graph comprising the second plurality of nodes connected by a second plurality of edges.