IP Library › Granted Patent US 12,530,577
Granted Patent B2
US 12,530,577 · App. 17/508,592 · Granted Jan 20, 2026

Training device, inferring device, training method, inferring method, and non-transitory computer readable medium

Inventors: Kaushalya Madhawa Pituwala Kankanamge (Tokyo, JP); Kosuke Nakago (Tokyo, JP); Katsuhiko Ishiguro (Tokyo, JP)
Assignee: Preferred Networks, Inc.
G06N3/08G06F18/214G06N3/045
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,530,577
App. No.
17/508,592
Granted
Jan 20, 2026
Kind
B2
Abstract

A training device comprises one or more memories and one or more processors. The one or more processors are configured to train a first converter that converts a first feature amount regarding a node of a graph and a second feature amount regarding a structure of the graph into a first latent value through conversion capable of defining inverse conversion, and a second converter that converts the second feature amount into a second latent value through conversion capable of defining inverse conversion, based on the first latent value and the second latent value.

Claims (80)

1 . A training device comprising:

one or more memories; and

one or more processors configured to:

convert, by using a first converter including a first neural network model, a feature regarding a node of a graph and a feature regarding a structure of the graph into a first latent value,

convert, by using a second converter including a second neural network model, the feature regarding the structure of the graph into a second latent value,

train, based on the first latent value and the second latent value, the first neural network model and the second neural network model,

wherein the graph is at least one of a graph of a compound, a graph of a circuit design, a graph of a transportation system, a graph of a network, a graph of architecture, a graph of linguistics, or a graph of a web,

wherein the first converter is a converter that is capable of defining inverse conversion,

wherein the second converter is a converter that is capable of defining inverse conversion,

wherein the one or more processors are further configured to define a first inverse converter and a second inverse converter, the first inverse converter being a converter that is configured to execute, by using the first neural network model, inverse conversion of the first converter, the second inverse converter being a converter that is configured to execute, by using the second neural network model, inverse conversion of the second converter, and

wherein the second inverse converter is configured to generate a feature regarding a structure of another graph and the first inverse converter is configured to generate a feature regarding a node of the another graph.

2 . The training device according to claim 1 , wherein

the one or more processors are configured to perform, based on nonlinear mapping, the conversion by the first converter and the conversion by the second converter.

3 . The training device according to claim 2 , wherein

the one or more processors are configured to execute the mapping based on NVP (Non-volume Preserving) mapping.

4 . The training device according to claim 1 , wherein

the one or more processors are configured to calculate the first latent value by applying multiple mappings on the feature regarding the node of the graph and the feature regarding the structure of the graph; and

the one or more processors are configured to calculate the second latent value by applying multiple mappings on the feature regarding the structure of the graph.

5 . The training device according to claim 1 , wherein

the graph is the graph of the compound,

the feature regarding the node of the graph is information regarding an atom of the compound, and

the feature regarding the structure of the graph is information regarding an adjacent state.

6 . The training device according to claim 1 , wherein

the feature regarding the node of the graph is a first tensor indicating a node feature of the graph; and

the feature regarding the structure of the graph is a second tensor indicating the structure of the graph.

7 . The training device according to claim 6 , wherein

the one or more processors are further configured to process acquired data regarding the graph into the first tensor and the second tensor, the second tensor including an adjacency matrix of the graph.

8 . The training device according to claim 1 , wherein

the one or more processors are configured to train the first neural network and the second neural network so that the first latent value and the second latent value follow a predetermined distribution.

9 . The training device according to claim 1 , wherein

the one or more processors train the first neural network and the second neural network based on a prior distribution with respect to the first latent value and the second latent value.

10 . An generation device comprising:

one or more memories; and

one or more processors configured to:

convert second data into a feature regarding a structure of a first graph by using a second inverse converter; and

convert first data into a feature regarding a node of the first graph by using a first inverse converter and the feature regarding the structure of the first graph,

wherein the first graph is at least one of a graph of a compound, a graph of a circuit design, a graph of a transportation system, a graph of a network, a graph of architecture, a graph of linguistics, or a graph of a web,

wherein the first inverse converter is defined based on a first converter capable of defining inverse conversion, the first converter including a first neural network model,

wherein the second inverse converter is defined based on a second converter capable of defining inverse conversion, the second converter including a second neural network model,

wherein the conversion executed by using the first inverse converter is inverse conversion of the first converter, and

wherein the conversion executed by using the second inverse converter is inverse conversion of the second converter.

11 . The generation device according to claim 10 , wherein

the first data is data obtained, by the one or more processors, by converting a feature regarding a node of a second graph and a feature regarding a structure of the second graph into a first latent value by using the first converter capable of defining inverse conversion; and

the second data is data obtained, by the one or more processors, by converting the feature regarding the structure of the second graph into a second latent value by using the second converter capable of defining inverse conversion.

12 . The generation device according to claim 11 , wherein

the first graph and the second graph are the same graph.

13 . The generation device according to claim 10 , wherein

the first data and the second data are generated based on random numbers.

14 . The generation device according to claim 10 , wherein

the first data and the second data are generated based on a predetermined distribution.

15 . The generation device according to claim 10 , wherein

the one or more processors are further configured to apply noise to at least one of the first data and the second data before converting the first data and the second data.

16 . The generation device according to claim 10 , wherein

the one or more processors are further configured to generate data of the first graph based on the feature regarding the node of the first graph and the feature regarding the structure of the first graph.

17 . The generation device according to claim 16 , wherein

the one or more processors are further configured to generate the first graph by a one-shot method.

18 . The generation device according to claim 10 , wherein

the feature regarding the node of the first graph is a first tensor indicating a node feature of the first graph; and

the feature regarding the structure of the first graph is a second tensor indicating the structure of the first graph.

19 . The generation device according to claim 10 , wherein

the first graph is the graph of the compound,

the feature regarding the node of the first graph includes information regarding an atom of the compound, and

the feature regarding the structure of the first graph includes information regarding an adjacent state.

20 . A training method comprising:

converting, by one or more processors, by using a first converter including a first neural network model, a feature regarding a node of a graph and a feature regarding a structure of the graph into a first latent value,

converting, by the one or more processors, by using a second converter including a second neural network model, the feature regarding the structure of the graph into a second latent value,

training, by the one or more processors, based on the first latent value and the second latent value, the first neural network model and the second neural network model,

wherein the graph is at least one of a graph of a compound, a graph of a circuit design, a graph of a transportation system, a graph of a network, a graph of architecture, a graph of linguistics, or a graph of a web,

wherein the first converter is a converter that is capable of defining inverse conversion,

wherein the second converter is a converter that is capable of defining inverse conversion,

wherein the method further comprises defining, by the one or more processors, a first inverse converter and a second inverse converter, the first inverse converter being a converter that is configured to execute, by using the first neural network model, inverse conversion of the first converter, the second inverse converter being a converter that is configured to execute, by using the second neural network model, inverse conversion of the second converter, and

wherein the second inverse converter is configured to generate a feature regarding a structure of another graph and the first inverse converter is configured to generate a feature regarding a node of the another graph.

21 . A method for generating graph comprising:

converting, by one or more processors, second data into a feature regarding a structure of a first graph by using a second inverse converter;

converting, by the one or more processors, first data into a feature regarding a node of the first graph by using a first inverse converter and the feature regarding the structure of the first graph,

wherein the first graph is at least one of a graph of a compound, a graph of a circuit design, a graph of a transportation system, a graph of a network, a graph of architecture, a graph of linguistics, or a graph of a web,

wherein the first inverse converter is defined based on a first converter capable of defining inverse conversion, the first converter including a first neural network model,

wherein the second inverse converter is defined based on a second converter capable of defining inverse conversion, the second converter including a second neural network model,

wherein the conversion executed by using the first inverse converter is inverse conversion of the first converter and,

wherein the conversion executed by using the second inverse converter is inverse conversion of the second converter.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Mar 21, 2022
From: PITUWALA KANKANAMGE, KAUSHALYA MADHAWA; NAKAGO, KOSUKE; ISHIGURO, KATSUHIKO
To: PREFERRED NETWORKS, INC.
Reel/Frame 059329/0710 →
Priority Claims (1)
JP 2019-082977 · Apr 24, 2019 · national
Continuity (2)
Continuation PCTJP2020003052 · Jan 28, 2020
Related Publication 20220044121A1 · Feb 10, 2022
References Cited (23)
US 20200110777A1 · Yin · 2020 [cited by examiner]
US 20200334495A1 · Al-Rfou · 2020 [cited by examiner]
US 20200401916A1 · Rolfe · 2020 [cited by examiner]
US 20210027864A1 · Plumbley · 2021 [cited by examiner]
US 20210081804A1 · Stojevic · 2021 [cited by examiner]
US 20220383992A1 · Triendl · 2022 [cited by examiner]
Jin et al “Learning Multimodal Graph-to-Graph Translation for Molecular Optimization”, dated Jan. 2019, and retrived from https://arxiv.org/abs/1812.01070v3 (Year: 2019). [cited by examiner]
Aviral Kumar et al., “GRevnet: Improving Graph Neural Nets with Reversible Computation”, Relational Representation Learning Workshop at 32nd Conference on Neural Information Processing Systems (NIPS 2018). [cited by applicant]
Diederik P. Kingma et al., “Auto-Encoding Variational Bayes”, arXiv:1312.6114v10 [stat.ML] May 1, 2014. [cited by applicant]
Diederik P. Kingma et al., “Glow: Generative Flow with Invertible 1×1 Convolutions”, arXiv:1807.03039v2 [stat.ML] Jul. 10, 2018. [cited by applicant]
Ian J. Goodfellow et al., “Generative Adversarial Nets”, arXiv:1406.2661v1 [stat.ML] Jun. 10, 2014. [cited by applicant]
Jiaxuan You et al., “Appendix for GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models”, Proceedings of the 35th International Conference on Machine Learning, Stockholm, Sweden, PMLR 80, 2018. [cited by applicant]
Jiaxuan You et al., “Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generation”, 32nd Conference on Neural Information Processing Systems (NeurIPS 2018) Montréal, Canada. [cited by applicant]
Jiaxuan You et al., “GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models”, Proceedings of the 35th International Conference on Machine Learning, Stockholm, Sweden, PMLR 80, 2018. [cited by applicant]
Laurent Dinh et al., “Density Estimation Using Real NVP”, arXiv:1605.08803v3 [cs.LG] Feb. 27, 2017. [cited by applicant]
Laurent Dinh et al., “NICE: Non-linear Independent Components Estimation”, arXiv:1410.8516v6 [cs.LG] Apr. 10, 2015. [cited by applicant]
Michael Schlichtkrull et al., “Modeling Relational Data with Graph Convolutional Networks”, arXiv:1703.06103v4 [stat.ML] Oct. 26, 2017. [cited by applicant]
Nicola De Cao et al., “MolGAN: An implicit generative model for small molecular graphs”, arXiv:1805.11973v1 [stat.ML] May 30, 2018. [cited by applicant]
Rafael Gomez-Bombarelli et al., “Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules”, ACS Central Science, 2018, 4, 268-276. [cited by applicant]
GraphRNN: Generating Realistic Graphs with Deep Auto-regressive Models http://web archive.org/web/20180711025304/http://proceedings.mlr.press/v80/you18a.html Jul. 11, 2018 Version, Abstract. [cited by applicant]
Prafulla Dhariwal et al., “Glow: Better Reversible Generative Models”, https://web.archive.org/web/20190311213439/https://openai.com/blog/glow/ Mar. 11, 2019 Version. [cited by applicant]
Wengong Jin et al., “Junction Tree Variational Autoencoder for Molecular Graph Generation”, The 35th International Conference on Machine Learning, Stockholm, Sweden, PMLR 80, 2018. [cited by applicant]
Wengong Jin et al., “Junction Tree Variational Autoencoder for Molecular Graph Generation”, The 35th International Conference on Machine Learning, Stockholm, Sweden, PMLR 80, 2018, arXiv: 1802.04364v4 [cs.LG] Mar. 29, 2… [cited by applicant]