# US 11,443,189 B2 · Hypernetwork training method and device, electronic device and storage medium > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 11,443,189 B2 - Document ID: US11443189B2 - Status: Granted - Kind: B2 - Application number: 16/828,427 - Filed: 2020-03-24 - Published: 2021-05-13 - Granted: 2022-09-13 - Earliest claimed: 2019-11-12 - Assignees: Beijing Xiaomi Intelligent Technology Co., Ltd. - Applicants: Beijing Xiaomi Intelligent Technology Co., Ltd. - Inventors: Xiangxiang Chu; Bo Zhang; Ruijun Xu; Bin Wang - Web page: https://patentgrep.com/patent/US11443189B2 ## CPC - G06N 3/08: Learning methods - G06N 3/04: Architecture, e.g. interconnection topology - G06N 3/045: Combinations of networks - G06N 3/0464: Convolutional networks [CNN, ConvNet] - G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections - G06N 3/09: Supervised learning - G06N 3/047: Probabilistic or stochastic networks - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract A hypernetwork training method includes: acquiring a multipath neural subnetwork based on a preconstructed initial hypernetwork; training the multipath neural subnetwork to update a weight parameter of each substructure in the multipath neural subnetwork; synchronizing the weight parameter of each substructure in the multipath neural subnetwork to the preconstructed initial hypernetwork; and determining whether the preconstructed initial hypernetwork converges, and if it is determined that the preconstructed initial hypernetwork does not converge, re-executing the acquiring, the training, the synchronizing, and the determining, to obtain a target hypernetwork. ## Claims ### Claim 1 A hypernetwork training method, comprising: acquiring a multipath neural subnetwork based on a preconstructed initial hypernetwork; training the multipath neural subnetwork to update a weight parameter of each substructure in the multipath neural subnetwork; synchronizing the weight parameter of each substructure in the multipath neural subnetwork to the preconstructed initial hypernetwork; and determining whether the preconstructed initial hypernetwork converges, and if it is determined that the preconstructed initial hypernetwork does not converge, re-executing the acquiring, the training, the synchronizing, and the determining, to obtain a target hypernetwork. ### Claim 2 (depends on 1) The hypernetwork training method of claim 1, wherein the acquiring the multipath neural subnetwork based on the preconstructed initial hypernetwork comprises: for each layer of the preconstructed initial hypernetwork, sampling at least one substructure in each layer, and adding an output end configured to add an output result of the at least one substructure to obtain a substructure portion comprising the at least one substructure and the output end; connecting each substructure portion in series according to a sequence of each layer to obtain an initial multipath neural subnetwork; and performing parameter sharing on a same substructure in the preconstructed initial hypernetwork and the initial multipath neural subnetwork to obtain a final multipath neural subnetwork. ### Claim 3 (depends on 2) The hypernetwork training method of claim 2, wherein the sampling the at least one substructure in each layer comprises: sequentially sampling the at least one substructure by a Bernoulli sampling method, wherein each substructure in each layer is sampled at an equal probability. ### Claim 4 (depends on 1) The hypernetwork training method of claim 1, wherein the training the multipath neural subnetwork comprises: acquiring a first number of training samples in a training set; and sequentially inputting the first number of training samples to the multipath neural subnetwork to train the multipath neural subnetwork until the first number of training samples are all used. ### Claim 5 (depends on 1) The hypernetwork training method of claim 1, further comprising: evaluating multiple multipath neural subnetworks, which are newly acquired, based on the target hypernetwork to obtain evaluation results; and determining a multipath neural subnetwork meeting a preset condition to apply the determined multipath neural subnetwork to a target scenario according to the evaluation results. ### Claim 6 (depends on 5) The hypernetwork training method of claim 5, wherein the evaluating the multiple multipath neural subnetworks, which are newly acquired, based on the target hypernetwork to obtain the evaluation results comprises: acquiring the multiple multipath neural subnetworks based on the target hypernetwork; sharing a weight parameter of each substructure in the target hypernetwork with each corresponding substructure in the multiple multipath neural subnetworks; and evaluating the multiple multipath neural subnetworks to obtain the evaluation results of the multiple multipath neural subnetworks. ### Claim 7 (depends on 5) The hypernetwork training method of claim 5, wherein the target scenario comprises at least one of image classification, target detection, semantic segmentation, text to speech, natural language translation, and speech enhancement. ### Claim 8 An electronic device, comprising: a processor; and a memory configured to store instructions executable by the processor, wherein the processor is configured to: acquire a multipath neural subnetwork based on a preconstructed initial hypernetwork; train the multipath neural subnetwork to update a weight parameter of each substructure in the multipath neural subnetwork; synchronize the weight parameter of each substructure in the multipath neural subnetwork to the preconstructed initial hypernetwork; and determine whether the preconstructed initial hypernetwork converges, and if it is determined that the preconstructed initial hypernetwork does not converge, re-execute the acquiring, the training, the synchronizing, and the determining, to obtain a target hypernetwork. ### Claim 9 (depends on 8) The electronic device of claim 8, wherein the processor is further configured to: for each layer of the preconstructed initial hypernetwork, sample at least one substructure in each layer and add an output end configured to add an output result of the at least one substructure to obtain a substructure portion comprising the at least one substructure and the output end; connect each substructure portion in series according to a sequence of each layer to obtain an initial multipath neural subnetwork; and perform parameter sharing on a same substructure in the preconstructed initial hypernetwork and the initial multipath neural subnetwork to obtain a final multipath neural subnetwork. ### Claim 10 (depends on 9) The electronic device of claim 9, wherein the processor is further configured to sequentially sample the at least one substructure by a Bernoulli sampling method, wherein each substructure in each layer is sampled at an equal probability. ### Claim 11 (depends on 8) The electronic device of claim 8, wherein the processor is further configured to: acquire a first number of training samples in a training set; and sequentially input the first number of training samples to the multipath neural subnetwork to train the multipath neural subnetwork until the first number of training samples are all used. ### Claim 12 (depends on 8) The electronic device of claim 8, wherein the processor is further configured to: evaluate multiple multipath neural subnetworks, which are newly acquired, based on the target hypernetwork to obtain evaluation results; and determine a multipath neural subnetwork meeting a preset condition, to apply the determined multipath neural subnetwork to a target scenario according to the evaluation results. ### Claim 13 (depends on 12) The electronic device of claim 12, wherein the processor is further configured to: acquire the multiple multipath neural subnetworks based on the target hypernetwork; share a weight parameter of each substructure in the target hypernetwork with each corresponding substructure in the multiple multipath neural subnetworks; and evaluate the multiple multipath neural subnetworks to obtain the evaluation results of the multiple multipath neural subnetworks. ### Claim 14 (depends on 12) The electronic device of claim 12, wherein the target scenario comprises at least one of image classification, target detection, semantic segmentation, text to speech, natural language translation, and speech enhancement. ### Claim 15 A non-transitory readable storage medium having stored thereon instructions that, when executed by a processor of a device, cause the device to perform a hypernetwork training method, the method comprising: acquiring a multipath neural subnetwork based on a preconstructed initial hypernetwork; training the multipath neural subnetwork to update a weight parameter of each substructure in the multipath neural subnetwork; synchronizing the weight parameter of each substructure in the multipath neural subnetwork to the preconstructed initial hypernetwork; and determining whether the preconstructed initial hypernetwork converges, and if it is determined that the preconstructed initial hypernetwork does not converge, re-executing the acquiring, the training, the synchronizing, and the determining, to obtain a target hypernetwork. ### Claim 16 (depends on 15) The non-transitory readable storage medium according to claim 15, wherein the acquiring the multipath neural subnetwork based on the preconstructed initial hypernetwork comprises: for each layer of the preconstructed initial hypernetwork, sampling at least one substructure in each layer and adding an output end configured to add an output result of the at least one substructure to obtain a substructure portion comprising the at least one substructure and the output end; connecting each substructure portion in series according to a sequence of each layer to obtain an initial multipath neural subnetwork; and performing parameter sharing on a same substructure in the preconstructed initial hypernetwork and the initial multipath neural subnetwork to obtain a final multipath neural subnetwork. ### Claim 17 (depends on 16) The non-transitory readable storage medium according to claim 16, wherein the sampling the at least one substructure in each layer comprises: sequentially sampling the at least one substructure by a Bernoulli sampling method, wherein each substructure in each layer is sampled at an equal probability. ### Claim 18 (depends on 15) The non-transitory readable storage medium according to claim 15, wherein the training the multipath neural subnetwork comprises: acquiring a first number of training samples in a training set; and sequentially inputting the first number of training samples to the multipath neural subnetwork to train the multipath neural subnetwork until the first number of training samples are all used. ### Claim 19 (depends on 15) The non-transitory readable storage medium according to claim 15, wherein the method further comprises: evaluating multiple multipath neural subnetworks, which are newly acquired, based on the target hypernetwork to obtain evaluation results; and determining a multipath neural subnetwork meeting a preset condition, to apply the determined multipath neural subnetwork to a target scenario according to the evaluation results. ### Claim 20 (depends on 19) The non-transitory readable storage medium according to claim 19, wherein the evaluating the multiple multipath neural subnetworks, which are newly acquired, based on the target hypernetwork to obtain the evaluation results comprises: acquiring the multiple multipath neural subnetworks based on the target hypernetwork; sharing a weight parameter of each substructure in the target hypernetwork with each corresponding substructure in the multiple multipath neural subnetworks; and evaluating the multiple multipath neural subnetworks to obtain the evaluation results of the multiple multipath neural subnetworks.