# US 12,217,185 B2 · Method and apparatus of increasing knowledge based on uncertainty in neural networks > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,217,185 B2 - Document ID: US12217185B2 - Status: Granted - Kind: B2 - Application number: 17/332,464 - Filed: 2021-05-27 - Published: 2021-12-02 - Granted: 2025-02-04 - Earliest claimed: 2020-05-29 - Assignees: Electronics and Telecommunications Research Institute - Applicants: Electronics and Telecommunications Research Institute - Inventors: Hyun Woo Kim; Jeon Gue Park; Hwa Jeon Song; Yoo Rhee Oh; Byung Hyun Yoo; Eui Sok Chung; Ran Han - Web page: https://patentgrep.com/patent/US12217185B2 ## CPC - G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections - G06F 40/35: Discourse or dialogue representation - G06N 3/042: Knowledge-based neural networks; Logical representations of neural networks - G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] - G06N 3/045: Combinations of networks - G06N 3/08: Learning methods - G06N 3/084: Backpropagation, e.g. using gradient descent - G06N 3/0895: Weakly supervised learning, e.g. semi-supervised or self-supervised learning - G06N 3/09: Supervised learning - G06N 3/091: Active learning - G06N 3/092: Reinforcement learning - G06N 5/022: Knowledge engineering; Knowledge acquisition - G06N 5/043: Distributed expert systems; Blackboards - G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) ## Abstract A knowledge increasing method includes calculating uncertainty of knowledge obtained from a neural network using an explicit memory, determining the insufficiency of the knowledge on the basis of the calculated uncertainty, obtaining additional data (learning data) for increasing insufficient knowledge, and training the neural network by using the additional data to autonomously increase knowledge. ## Claims ### Claim 1 A method of increasing knowledge based on uncertainty of a neural network in an artificial intelligence (AI) agent, programmed and executed by a computing device to include a knowledge generating neural network for generating knowledge about a user query and a dialogue generating neural network for generating an answer to the user query on the basis of the knowledge, the method comprising: converting each of a current user query, previous knowledge generated by the knowledge generating neural network, and a previous answer, generated by the dialogue generating neural network, into an embedding vector value having a specific dimension; calculating a plurality of output values corresponding to the embedding vector value by using the knowledge generating neural network with a dropout technique applied thereto and calculating uncertainty of the knowledge generating neural network on the basis of the calculated plurality of output values; and performing linear conversion on the calculated uncertainty to calculate a compensation value, obtain learning data having the calculated compensation value from a database built through unsupervised learning, and training the knowledge generating neural network and the dialogue generating neural network on the basis of the obtained learning data to increase knowledge. ### Claim 2 (depends on 1) The method of claim 1, wherein the calculating of the uncertainty comprises: deleting some neurons included in the knowledge generating neural network on the basis of the dropout technique; obtaining the plurality of output values, corresponding to number of methods of deleting the some neurons, from the knowledge generating neural network; and calculating the uncertainty on the basis of an average and a variance of the plurality of output values. ### Claim 3 (depends on 1) The method of claim 1, wherein the calculating of the uncertainty comprises calculating uncertainty, occurring in a process of modeling data, and uncertainty occurring due to noise of the data. ### Claim 4 (depends on 1) The method of claim 1, wherein the calculating of the uncertainty comprises: calculating an average of the plurality of output values corresponding to the embedding vector value by using the knowledge generating neural network with the dropout technique applied thereto; calculating first uncertainty occurring in a process of modeling data on the basis of the calculated average; calculating a variance of the plurality of output values; calculating second uncertainty caused by noise of the data on the basis of the calculated variance; and summating the first uncertainty and the second uncertainty to calculate uncertainty of the knowledge generating neural network. ### Claim 5 (depends on 1) The method of claim 1, wherein the increasing of the knowledge comprises: determining the insufficiency or not of the knowledge about a user demand matter corresponding to the embedding vector value on the basis of the calculated uncertainty; and when it is determined that the knowledge is insufficient, searching for the database to obtain the learning data. ### Claim 6 (depends on 5) The method of claim 5, wherein the determining of the insufficiency or not of the knowledge comprises comparing the calculated uncertainty with a predetermined threshold value to determine the insufficiency or not of the knowledge on the basis of a result of the comparison. ### Claim 7 (depends on 5) The method of claim 5, wherein the generating of the learning data comprises: performing linear conversion on the calculated uncertainty to calculate a compensation value; and obtaining learning data including the calculated compensation value from the database by using the embedding vector value as a query value. ### Claim 8 (depends on 1) The method of claim 1, wherein the increasing of the knowledge comprises training the knowledge generating neural network and the dialogue generating neural network through unsupervised learning. ### Claim 9 (depends on 1) The method of claim 1, wherein the AI agent further comprises a value estimation neural network programmed to generate a parameter for training the knowledge generating neural network and the dialogue generating neural network through reinforcement learning on the basis of the knowledge and an accuracy of the answer, and the increasing of the knowledge comprises further training the value estimation neural network on the basis of the generated learning data. ### Claim 10 A knowledge increasing apparatus comprising: a storage configured to store an artificial intelligence (AI) agent programmed to include a knowledge generating neural network for generating knowledge about a user query and a dialogue generating neural network for generating an answer to the user query on the basis of the knowledge; and a processor configured to execute the AI agent stored in the storage, wherein the AI agent comprises: a preprocessor that converts each of a current user query, previous knowledge generated by the knowledge generating neural network, and a previous answer, generated by the dialogue generating neural network, into an embedding vector value having a specific dimension; a calculator that calculates a plurality of output values corresponding to the embedding vector value by using the knowledge generating neural network with a dropout technique applied thereto and to calculate uncertainty of the knowledge generating neural network on the basis of the calculated plurality of output values; and a sampler that linearly converts on the calculated uncertainty to calculate a compensation value, obtain learning data having the calculated compensation value from a database built through unsupervised learning, and train the knowledge generating neural network and the dialogue generating neural network on the basis of the obtained learning data to increase knowledge. ### Claim 11 (depends on 10) The knowledge increasing apparatus of claim 10, wherein the calculator calculates first uncertainty occurring in a process of modeling data on the basis of an average of the plurality of output values, calculates second uncertainty caused by noise of the data on the basis of a variance of the plurality of output values, and summates the first uncertainty and the second uncertainty to calculate uncertainty of an output of the knowledge generating neural network. ### Claim 12 (depends on 10) The knowledge increasing apparatus of claim 10, wherein the sampler compares the calculated uncertainty with a predetermined threshold value to determine the insufficiency or not of the knowledge on the basis of a result of the comparison, and when it is determined that the knowledge is insufficient, searches for the database to obtain the learning data. ### Claim 13 (depends on 10) The knowledge increasing apparatus of claim 10, wherein the sampler obtains the learning data corresponding to insufficient knowledge from the database by using the embedding vector value as a query value. ### Claim 14 (depends on 10) The knowledge increasing apparatus of claim 10, wherein the sampler trains the knowledge generating neural network and the dialogue generating neural network through unsupervised learning to increase knowledge. ### Claim 15 (depends on 10) The knowledge increasing apparatus of claim 10, wherein the AI agent further comprises a value estimation neural network programmed to generate a parameter for training the knowledge generating neural network and the dialogue generating neural network through reinforcement learning on the basis of the knowledge and an accuracy of the answer, and the sampler further trains the value estimation neural network on the basis of the generated learning data.