# US 12,602,585 B2 · Methods and systems for determining a representative input data set for post-training quantization of artificial neural networks > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,602,585 B2 - Document ID: US12602585B2 - Status: Granted - Kind: B2 - Application number: 18/129,174 - Filed: 2023-03-31 - Published: 2023-10-12 - Granted: 2026-04-14 - Earliest claimed: 2022-04-07 - Assignees: Aptiv Technologies AG - Applicants: Aptiv Technologies AG - Inventors: Robert Lajos Bucs; Miguel Angel Aguilar; Soeren Kreinberg - Web page: https://patentgrep.com/patent/US12602585B2 ## CPC - G06N 3/063: using electronic means - G06N 3/08: Learning methods - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract A computer implemented method for determining a representative input data set for post-training quantization of artificial neural networks comprises: acquiring an input data set; processing the input data set to extract one or more parameters associated with the input data set; setting clipping limits comprising a lower clipping limit and an upper clipping limit; adapting the upper and lower clipping limits by computing (656) a quantization error based on a numerical difference between the input data set and a quantized and a dequantized input data set, wherein the error is computed based on an error metric which is based on user input; and determining a representative input data set as a subset of the input data set whose elements fit within the clipping limits; wherein the representative input data set is to be used for quantization calibration of the post-training quantization. ## Claims ### Claim 1 Computer implemented method for determining a representative input data set for post-training quantization of artificial neural networks, the method comprising of the following steps: acquiring an input data set; processing the input data set to extract one or more parameters associated with the input data set, wherein metrics applied to the input data set to extract the one or more parameters are based on user input; based on the extracted one or more parameters, setting initial values for clipping limits comprising a lower clipping limit and an upper clipping limit, wherein the lower clipping limit and the upper clipping limit define threshold values for limiting a value range of the input data set; adapting the upper and lower clipping limits by computing a quantization error based on a numerical difference between the input data set and a quantized and a dequantized input data set, wherein the error is computed based on an error metric which is based on user input; and determining a representative input data set as a subset of the input data set whose elements fit within the clipping limits; wherein the representative input data set is to be used for quantization calibration of the post-training quantization. ### Claim 2 (depends on 1) The method of claim 1, wherein the upper and lower clipping limits are adapted by iteratively carrying out: quantizing the input data set by mapping values of the input data set from a set of higher numerical precision to values of lower numerical precision based on the one or more extracted parameters and the clipping limits; dequantizing the quantized input data set by mapping values of the quantized input data set from lower numerical precision back to higher numerical precision based on the one or more extracted parameters and the clipping limits; computing the quantization error based on the numerical difference between the input data set and the quantized and dequantized input data set; and adapting the clipping limits based on the computed quantization error. ### Claim 3 (depends on 1) The method of claim 1, wherein the clipping limits are adapted until a difference between changed clipping limits from one iteration to another iteration are below a pre-determined threshold. ### Claim 4 (depends on 3) The method of claim 3, wherein the threshold is based on the input data set, preferably one or more parameters defined by the input data set; and wherein the threshold is based on user-input. ### Claim 5 (depends on 1) The method of claim 1, wherein the user input is provided in a configuration file, preferably a textual configuration file. ### Claim 6 (depends on 1) The method of claim 1, wherein the clipping limits define a lower threshold and an upper threshold which are used to delimit the quantization of the input data set. ### Claim 7 (depends on 1) The method of claim 1, further comprising: determining a scale, wherein the scale is used for quantization and dequantization. ### Claim 8 (depends on 1) The method of claim 1, wherein determining the representative input data set comprises determining an overlap between an interval defined by the clipping limits and the input data set. ### Claim 9 (depends on 1) The method of claim 1, wherein the quantization error is determined based on a user-defined error function; and/or wherein the quantization error is determined based on a mean squared error and/or a mean average error and/or a peak signal to noise ratio and/or the a Kullback-Leibler divergence metric. ### Claim 10 (depends on 1) The method of claim 1, wherein quantization calibration comprises calibrating parameters for a network quantization of artificial neural networks. ### Claim 11 (depends on 1) The method of claim 1, wherein the artificial neural network is to be deployed on a resource-constrained embedded system. ### Claim 12 (depends on 11) The method of claim 11, wherein the embedded system is a mobile computing device, a mobile phone, a tablet computing device, an automotive compute platform, or an edge device. ### Claim 13 (depends on 1) A Computer system comprising a plurality of computer hardware components configured to carry out steps of the method of claim 1. ### Claim 14 (depends on 1) Non-transitory computer readable medium comprising instructions for carrying out the method of claim 1.