System and method using separable transfer learning-based artificial neural network

A computing system using machine learning-based artificial intelligence, the computing system comprising:

• at least one processor,
• a first artificial neural network configured to be pre-trained on a first phase task based on a first dataset;
• and
• a second artificial neural network configured to be pre-trained on a second phase task based on a second dataset,
• wherein the at least one processor is configured to:
• determine the second artificial neural network as a target artificial neural network configured to be trained on a target task based on a target dataset after a first pre-training of the first artificial neural network on the first phase task based on the first dataset and a second pre-training of the second artificial neural network on the second phase task based on the second dataset;
• control the first pre-training of the first artificial neural network and the second pre-training of the second artificial neural network to be performed independently of each other;
• set a plurality of candidates for phase layer region information including a boundary, based on which a first target phase layer and a second target phase layer are distinguished from each other in the target artificial neural network;
• control a transfer of, for each of the plurality of candidates, a parameter of a first phase layer of the first artificial neural network to the first target phase layer of the target artificial neural network after the first pre-training of the first artificial neural network and the second pre-training of the second artificial neural network, wherein the first phase layer is determined based on the first target phase layer;
• control, for each of the plurality of candidates, the target artificial neural network to perform fine-tuning training on the target task after the transfer of the parameter of the first phase layer of the first artificial neural network to the target artificial neural network;
• and
• determine the phase layer region information from among the plurality of candidates based on respective prediction accuracy for the target task provided by the target artificial neural network performing the transfer and the fine-tuning training for each of the plurality of candidates,
• wherein the target dataset, the first dataset, and the second dataset are selected based on personal characteristic information and schedule and lifestyle data of a user,
• wherein the at least one processor is configured to infer a time point, at which a need for the user to wear a smart wearable device is recognized to be high, based on results of learning for the schedule and lifestyle data of the user, and
• wherein the smart wearable device comprises at least one selected from the group consisting of a smart hip protection for preventing an injury to the user, and a smart healthcare device capable of detecting biosignals.

An artificial neural network training method executed in a computing system including a machine learning-based artificial neural network, the artificial neural network training method comprising:

• controlling, by at least one processor in a computing system, a first pre-training of a first artificial neural network so that the first artificial neural network is pre-trained on a first phase task based on a first dataset;
• controlling, by the at least one processor, a second pre-training of a second artificial neural network so that the second artificial neural network is pre-trained on a second phase task based on a second dataset, wherein the first pre-training of the first artificial neural network on the first phase task based on the first dataset and the second pre-training of the second artificial neural network on the second phase task based on the second dataset are controlled to be performed independently of each other;
• setting, by the at least one processor, a plurality of candidates for phase layer region information including a boundary, based on which a first target phase layer and a second target phase layer are distinguished from each other in a target artificial neural network;
• controlling a first transfer of, by the at least one processor, for each of the plurality of candidates, a parameter of a first phase layer of the pre-trained first artificial neural network to a first target phase layer of the target artificial neural network, wherein the first phase layer is determined based on the first target phase layer;
• controlling a second transfer of, by the at least one processor, for each of the plurality of candidates, a parameter of a second phase layer of the pre-trained second artificial neural network to a second target phase layer of the target artificial neural network, wherein the second phase layer is determined based on the second target phase layer;
• controlling, by the at least one processor, for each of the plurality of candidates, fine-tuning training of the target artificial neural network so that the target artificial neural network performs fine-tuning training on a target task based on a target dataset;
• determining, by the at least one processor, the phase layer region information from among the plurality of candidates based on respective prediction accuracy for the target task provided by the target artificial neural network performing the first transfer, the second transfer, and the fine-tuning training for each of the plurality of candidates;
• and
• providing, by the at least one processor, the target artificial neural network, fine-tuned and trained based on the determined phase layer region information, as a final target artificial neural network to a user,
• wherein the target dataset, the first dataset, and the second dataset are selected based on personal characteristic information and schedule and lifestyle data of a user,
• wherein the at least one processor is configured to infer a time point, at which a need for the user to wear a smart wearable device is recognized to be high, based on results of learning for the schedule and lifestyle data of the user, and
• wherein the smart wearable device comprises at least one selected from the group consisting of a smart hip protection for preventing an injury to the user, and a smart healthcare device capable of detecting biosignals.