# US 11,113,596 B2 · Select one of plurality of neural networks > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 11,113,596 B2 - Document ID: US11113596B2 - Status: Granted - Kind: B2 - Application number: 15/574,083 - Filed: 2015-05-22 - Published: 2018-10-18 - Granted: 2021-09-07 - Earliest claimed: 2015-05-22 - Assignees: Longsand Limited - Applicants: Longsand Limited - Inventors: David Pye; Christopher Waple - Web page: https://patentgrep.com/patent/US11113596B2 ## CPC - G06N 3/105: {Shells for specifying net layout} - G06N 3/0464: Convolutional networks [CNN, ConvNet] - G06N 3/10: Interfaces, programming languages or software development kits, e.g. for simulating neural networks - G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS ## Abstract Data is input to one of a plurality of neural networks. Each of the plurality of neural networks is to be of a different size. A propagation time is determined for the inputted data. The propagation time relates to a time for the inputted data to propagate through one of the plurality of neural networks. One of the plurality of neural networks is selected based on the propagation time. ## Claims ### Claim 1 A device, comprising: an input unit that provides input data to one of a plurality of neural networks, each of the plurality of neural networks to be of a different size; a propagation unit that determines a propagation time for the input data, the propagation time to relate to a time for the input data to propagate through one of the plurality of neural networks; and a select unit that selects the one of the plurality of neural networks based on a comparison of the propagation time to a second time that comprises an amount of time used to transcribe the input data after the input data has propagated through the one of the plurality of neural networks. ### Claim 2 (depends on 1) The device of claim 1, wherein, the propagation unit continuously updates the propagation time to include a sum of propagation times for all of the input data that has propagated through the one of the plurality of neural networks. ### Claim 3 (depends on 2) The device of claim 2, wherein, the select unit calculates a ratio based on the propagation time and the second time, and the select unit selects the one of the plurality of neural networks based on a comparison of the calculated ratio to one or more target values. ### Claim 4 (depends on 3) The device of claim 3, wherein, the select unit switches to a smaller neural network of the plurality of neural networks if the calculated ratio is greater than one or more of the target values, the select unit switches to a larger neural network of the plurality of neural networks if the calculated ratio is less than one or more of the target values, and the select unit switches to a midsize neural network having a size between the smaller and larger neural networks if the calculated ratio is between at least two of the one or more of the target values. ### Claim 5 (depends on 4) The device of claim 4, wherein, the select unit initially selects the midsize neural network if the propagation time has not been determined by the propagation unit, and the select unit selects a largest neural network of the plurality of neural networks if one or more of the target values have not been determined. ### Claim 6 (depends on 4) The device of claim 4, wherein, the input data is divided into a plurality of units of data, and the input unit inputs the plurality of units of data into the selected one of the plurality of neural networks on a unit-by-unit basis. ### Claim 7 (depends on 6) The device of claim 6, wherein, the propagation unit measures a time for each of the plurality of units of data to propagate through one of the neural networks, the propagation unit continuously adds the measured times for the plurality of units of data to propagate through one of the neural networks to the propagation time, the propagation unit continuously updates a search time to include a sum of second times for the units of data that have propagated through one of the neural networks. ### Claim 8 (depends on 1) The device of claim 1, wherein, the plurality of neural networks output a plurality of posterior probabilities in relation to the input data, and each of the posterior probabilities are to relate to a probability for a different corresponding output for the input data. ### Claim 9 (depends on 8) The device of claim 8, wherein, a larger neural network of the plurality of neural networks more accurately determines the posterior probabilities than a smaller neural network of the plurality of neural networks, and the smaller neural network determines the posterior probabilities in less time than the larger neural network. ### Claim 10 (depends on 1) The device of claim 1, wherein, the plurality of neural networks have a same number of input nodes and a same number of output nodes, and the plurality of neural networks have a different number of hidden layer nodes. ### Claim 11 (depends on 1) The device of claim 1, wherein, the input data includes at least one of an image and audio, the plurality of neural networks aid in transcribing the at least one of the audio and image to text, and the plurality of neural networks comprise at least one of deep and convolutional neural networks. ### Claim 12 A method, comprising: providing input data to one of a plurality of neural networks, each of the plurality of neural networks being a different size; measuring a propagation time for the input data that has propagated through one of the plurality of neural networks; determining a second time related to an amount of time used to transcribe the input data after the input data has propagated through the one of the plurality of neural networks; calculating a ratio based on the propagation time and the second time; and selecting the one of the plurality of neural networks based on a comparison of the calculated ratio to a target value. ### Claim 13 (depends on 12) The method of claim 12, wherein, the calculating comprises continuously updating the ratio, and the selecting comprises continuously selecting between the plurality of neural networks based on the comparison. ### Claim 14 (depends on 12) The method of claim 12, wherein, the plurality of neural networks have a same number of input nodes and a same number of output nodes, and the plurality of neural networks have a different number of hidden layer nodes. ### Claim 15 (depends on 12) The method of claim 12, wherein, the input data includes at least one of an image and audio, the plurality of neural networks aid in transcribing the at least one of the audio and image to text, and the plurality of neural networks comprise at least one of deep and convolutional neural networks. ### Claim 16 (depends on 12) The method of claim 12, wherein, the plurality of neural networks output a plurality of posterior probabilities in relation to the input data, and each of the posterior probabilities relate to a probability for a different corresponding output for the input data. ### Claim 17 (depends on 12) The method of claim 12, wherein the input data is divided into a plurality of units of data that are provided into the selected one of the plurality of neural networks on a unit-by-unit basis. ### Claim 18 A non-transitory computer-readable storage medium storing instructions that, if executed by a processor of a device, cause the processor to: measure a propagation time for data to propagate through one of a plurality of neural networks, where each of the plurality of neural networks is a different size; determine a second time that comprises an amount of time used to transcribe the data after the data has propagated through one of the plurality of neural networks; and select the one of the plurality of neural networks based on the propagation time and the second time. ### Claim 19 (depends on 18) The non-transitory computer-readable storage medium of claim 18, wherein, a ratio is determined based on the propagation time and the second time, and the one of the plurality of neural networks is selected based on a comparison of the ratio to a target value. ### Claim 20 (depends on 18) The non-transitory computer-readable storage medium of claim 18, wherein, the data includes at least one of an image and audio, and the plurality of neural networks aid in transcribing at least one of the image and audio to text.