# US 10,872,300 B2 · Artificial intelligent systems and methods for using a structurally simpler learner model to mimic behaviors of a structurally more complicated reference model > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 10,872,300 B2 - Document ID: US10872300B2 - Status: Granted - Kind: B2 - Application number: 16/930,341 - Filed: 2020-07-16 - Published: 2020-11-05 - Granted: 2020-12-22 - Earliest claimed: 2018-11-06 - Assignees: Beijing Didi Infinity Technology and Development Co., Ltd. - Applicants: Beijing Didi Infinity Technology and Development Co., Ltd. - Inventors: Yuan Zhao - Web page: https://patentgrep.com/patent/US10872300B2 ## CPC - G06N 20/00: Machine learning - G06F 18/22: Matching criteria, e.g. proximity measures - G06N 5/046: Forward inferencing; Production systems - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) ## Abstract A method for using a structurally more complicated reference model to train a structurally simpler learner model includes: obtaining a trained reference model at least including N reference blocks and a learner model at least including N learner blocks respectively corresponding to the N reference blocks; training the learner model by conducting an iterative operation; determining whether the learner model is convergent; and in response to that the learner model is convergent, stopping the iterative operation to assign the learner model as a trained learner model. The iterative operation includes inputting a sample data set into the trained reference model and the learner model; for each of the N learner blocks: determining a distance between a learner vector of the learner block and a reference vector of the reference block, and updating parameters in the learner block based on the determined distance. ## Claims ### Claim 1 An artificial intelligent system for using a structurally simpler learner model to mimic behaviors of a structurally more complicated reference model, comprising: at least one storage medium including a set of instructions for training a learner model; and at least one processor in communication with the storage medium, wherein when executing the set of instructions, the at least one processor is directed to: obtain a trained reference model at least including N reference blocks, wherein N is an integer greater than 2; obtain a learner model at least including N learner blocks respectively corresponding to the N reference blocks of the trained reference model; train the learner model into a trained learner model by conducting an iterative operation by: inputting a sample data set into the trained reference model and the learner model; for each of the N learner blocks: determining a reference vector based on an output of the reference block corresponding to the learner block, determining a learner vector based on an output of the learner block, wherein a length of the reference vector equals a length of the learner vector, determining a distance between the reference vector and the learner vector to obtain a learning signal, and updating parameters in the learner block by inputting the distance into the learner block as a learning signal; determine whether the learner model is convergent; and in response to a determination that the learner model is convergent, stop the iterative operation to assign the learner model as the trained learner model. ### Claim 2 (depends on 1) The system of claim 1, wherein each of the N reference blocks includes at least one reference layer, and each of the N learner blocks includes at least one learner layers. ### Claim 3 (depends on 2) The system of claim 2, wherein a total number of learner blocks in the learner model is less than a total number of reference blocks in the trained reference model; a total number of learner layers in the learner model is less than a total number of reference layers in the trained reference model; and a total number of nodes in each learner layer is less than a total number of nodes in each reference layer. ### Claim 4 (depends on 1) The system of claim 1, wherein the at least one processor is further directed to: map an output of each reference block into a reference vector including a predetermined length; and map an output of each learner block into a learner vector including the predetermined length. ### Claim 5 (depends on 1) The system of claim 1, wherein the distance between the learner vector and the reference vector includes at least one of: a Cross Entropy, a Mutual Information, a Kullback-Leibler Divergence, an Euclidean Distance, or an Edit Distance. ### Claim 6 (depends on 1) The system of claim 1, wherein during a current iterative operation to determine whether the learner model is convergent, the at least one processor is further directed to: determine a difference value between parameters of a current learner model in the current iterative operation and parameters of a last learner model in a last iterative operation; determine whether the difference value is less than a difference threshold; and in response to a determination that the difference value is less than a difference threshold, determine that the learner model is convergent. ### Claim 7 (depends on 1) The system of claim 1, wherein the at least one processor is further directed to: train a plurality of learner models to obtain a plurality of trained learner models, wherein each learner model in the plurality of learner models is different from one another; and test the plurality of trained learner models to obtain a final trained learner model. ### Claim 8 (depends on 7) The system of claim 7, wherein to test the plurality of trained learner models, the at least one processor is further directed to: for each of the plurality of trained learner models, determine an output difference between the trained learner model and the trained reference model using a test dataset as inputs of the trained learner model and the trained reference model; and determine a final trained learner model from the plurality of trained learner models based on the plurality of output differences. ### Claim 9 (depends on 1) The system of claim 1, wherein the parameters in each learner block are updated based on the corresponding learning signal and a Gradient Descent algorithm. ### Claim 10 An artificial intelligent method for using a structurally simpler learner model to mimic behaviors of a structurally more complicated reference model, comprising: obtaining a trained reference model at least including N reference blocks, wherein N is an integer greater than 2; obtaining a learner model at least including N learner blocks respectively corresponding to the N reference blocks of the trained reference model; training the learner model into a trained learner model by conducting an iterative operation by: inputting a sample data set into the trained reference model and the learner model; for each of the N learner blocks: determining a reference vector based on an output of the reference block corresponding to the learner block, determining a learner vector based on an output of the learner block, wherein a length of the reference vector equals a length of the learner vector, determining a distance between the reference vector and the learner vector to obtain a learning signal, and updating parameters in the learner block by inputting the distance into the learner block as a learning signal; determining whether the learner model is convergent; and in response to a determination that the learner model is convergent, stopping the iterative operation to assign the learner model as the trained learner model. ### Claim 11 (depends on 10) The method of claim 10, wherein each of the N reference blocks includes at least one reference layer, and each of the N learner blocks includes at least one learner layers. ### Claim 12 (depends on 11) The method of claim 11, wherein a total number of learner blocks in the learner model is less than a total number of reference blocks in the trained reference model; a total number of learner layers in the learner model is less than a total number of reference layers in the trained reference model; and a total number of nodes in each learner layer is less than a total number of nodes in each reference layer. ### Claim 13 (depends on 10) The method of claim 10, further comprising: mapping an output of each reference block into a reference vector including a predetermined length; and mapping an output of each learner block into a learner vector including the predetermined length. ### Claim 14 (depends on 10) The method of claim 10, wherein the distance between the learner vector and the reference vector includes at least one of: a Cross Entropy, a Mutual Information, a Kullback-Leibler Divergence, an Euclidean Distance, or an Edit Distance. ### Claim 15 (depends on 10) The method of claim 10, wherein during a current iterative operation, the determining whether the learner model is convergent, further includes: determining a difference value between parameters of a current learner model in the current iterative operation and parameters of a last learner model in a last iterative operation; determining whether the difference value is less than a difference threshold; and in response to a determination that the difference value is less than a difference threshold, determining that the learner model is convergent. ### Claim 16 (depends on 10) The method of claim 10, further comprising: training a plurality of learner models to obtain a plurality of trained learner models, wherein each learner model in the plurality of learner models is different from one another; and testing the plurality of trained learner models to obtain a final trained learner model. ### Claim 17 (depends on 16) The method of claim 16, wherein the testing of the plurality of trained learner models further includes: for each of the plurality of trained learner models, determining an output difference between the trained learner model and the trained reference model using a test dataset as inputs of the trained learner model and the trained reference model; and determining a final trained learner model from the plurality of trained learner models based on the plurality of output differences. ### Claim 18 (depends on 10) The method of claim 10, wherein the parameters in each learner block are updated based on the corresponding learning signal and a Gradient Descent algorithm. ### Claim 19 A non-transitory computer readable medium, comprising at least one set of artificially intelligent instructions compatible for using a structurally simpler learner model to mimic behaviors of a structurally more complicated reference model, wherein when executed by at least one processor of one or more electronic device, the at least one set of instructions directs the at least one processor to: obtain a trained reference model at least including N reference blocks, wherein N is an integer greater than 2; obtain a learner model at least including N learner blocks respectively corresponding to the N reference blocks of the trained reference model; train the learner model into a trained learner model by conducting an iterative operation by: inputting a sample data set into the trained reference model and the learner model; for each of the N learner blocks: determining a reference vector based on an output of the reference block corresponding to the learner block, determining a learner vector based on an output of the learner block, wherein a length of the reference vector equals a length of the learner vector, determining a distance between the reference vector and the learner vector to obtain a learning signal, and updating parameters in the learner block by inputting the distance into the learner block as a learning signal; determine whether the learner model is convergent; and in response to a determination that the learner model is convergent, stop the iterative operation to assign the learner model as the trained learner model. ### Claim 20 (depends on 19) The non-transitory computer readable medium of claim 19, wherein each of the N reference blocks includes at least one reference layer, and each of the N learner blocks includes at least one learner layers.