Training a refined machine learning model

By adjusting the temperature value of the softmax layer and integrating the full machine learning model with specialized models, a refined machine learning model was trained, solving the problems of deployment and computational resource redundancy, and achieving a refined model with high efficiency and classification accuracy.

CN115618941BActive Publication Date: 2026-11-17GOOGLE LLC
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202211258471.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2014-06-06
Filing Date
2015-06-08
Publication Date
2026-11-17
Estimated Expiration
2035-06-08

AI Technical Summary

Technical Problem

Existing machine learning models are redundant in terms of deployment and computing resources, and it is difficult to match the classification accuracy of such cumbersome models.

Method used

By using a temperature-adjusted softmax layer, and combining an ensemble model of full machine learning and specialized machine learning models, a refined machine learning model is trained to generate models that are easier to deploy and require fewer computational resources, while maintaining classification accuracy.

Benefits of technology

It achieves a refined machine learning model that is similar in classification accuracy to the cumbersome model, but is easier to deploy and requires less computational resources, making it suitable for classification tasks of images and text fragments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115618941B_ABST
    Figure CN115618941B_ABST
Patent Text Reader

Abstract

A method, system, and apparatus for training a refined machine learning model, comprising a computer program encoded on a computer storage medium. One method includes training a verbose machine learning model, wherein the verbose machine learning model is configured to receive inputs and generate corresponding scores for each of a plurality of categories; and training a refined machine learning model on a plurality of training inputs, wherein the refined machine learning model is also configured to receive inputs and generate scores for the plurality of categories, comprising: processing each training input with the verbose machine learning model to generate a verbose target soft output for that training input; and training the refined machine learning model to generate a soft output for each training input that matches the verbose target soft output for that training input.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Case Analysis

[0002] This application is a divisional application of Chinese invention patent application 201510435904.1, filed on June 8, 2015. Technical Field

[0003] This manual pertains to training machine learning models. Background Technology

[0004] Machine learning models receive input and generate output based on the received input and the model's parameter values. For example, a machine learning model can receive an image and generate a score for each category set, where the score for a given category represents the probability that the image contains an image of an object belonging to that category.

[0005] Machine learning models can consist of a single level of linear or non-linear operations, or they can be deep networks, which are machine learning models consisting of multiple levels, one or more of which may be non-linear operations. An example of a deep network is a neural network with one or more hidden layers. Summary of the Invention

[0006] In summary, this specification describes techniques for training refined machine learning models using complex machine learning models.

[0007] Specific embodiments of the subject matter described in this specification may be implemented to achieve one or more of the following advantages. A refined machine learning model is easier to deploy than a verbose machine learning model, i.e., because it requires less computation, storage, or both, to generate output at runtime, even when the refined model can be efficiently trained using a verbose neural network. Once trained using a verbose machine learning model, the refined model can generate outputs with an accuracy not significantly lower than that generated by the verbose model, despite being easier to deploy or using fewer computational resources.

[0008] An ensemble model that includes one or more full machine learning models and one or more specialized machine learning models can generate more accurate scores for classifying received inputs. In particular, by including specialized machine learning models in the ensemble model, scores can be generated more accurately for categories that are frequently predicted to be together or confused by the full machine learning model.

[0009] Details of one or more embodiments of the subject matter of this specification are set forth in the accompanying drawings and the following description. Other features, aspects, and advantages of the subject matter will become apparent from this description, the drawings, and the claims. Attached Figure Description

[0010] Figure 1 An example of a refined machine learning model training system is shown.

[0011] Figure 2 This is a flowchart illustrating an example process of training a refined machine learning model using a previously trained, verbose machine learning model.

[0012] Figure 3 An example machine learning model system is shown.

[0013] Figure 4 This is a flowchart illustrating an example process of input processing using an ensemble machine learning model that includes one or more full machine learning models and one or more specialized machine learning models.

[0014] The same reference numerals and names in different figures represent the same elements. Detailed Implementation

[0015] Figure 1 This is a block diagram of an example refined machine learning model training system 100 for training a refined machine learning model 120. The refined machine learning model training system 100 is an example of a system implemented as a computer program on one or more computers in one or more locations, wherein the systems, components and techniques described below are implemented.

[0016] The refined machine learning model training system 100 uses the trained, cumbersome machine learning model 110 to train the refined machine learning model 120. Generally, a machine learning model receives input and generates output based on the received input and the model's parameter values.

[0017] Specifically, both the refined machine learning model 120 and the trained verbose machine learning model 110 are machine learning models configured to receive input and process the received input to generate corresponding scores for each category in a predetermined set of categories. Generally, the refined machine learning model 120 is a model with a different architecture than the verbose machine learning model 110, making it easier to deploy relative to the verbose machine learning model 110, for example, because the refined machine learning model 120 requires less computation, storage, or both than the verbose machine learning model 110 to generate output at runtime. For example, the refined machine learning model 120 may have fewer layers, fewer parameters, or both than the verbose machine learning model 110.

[0018] The trained bloat machine learning model 110 has been trained on a set of training inputs using conventional machine learning training techniques to determine the trained parameter values ​​of the bloat machine learning model 110. In particular, the trained bloat machine learning model 110 has been trained such that the score generated by the trained bloat machine learning model 110 for a given input and a given classification represents the probability that the classification is the exact classification of the input.

[0019] For example, if the input to the verbose machine learning model 110 is an image, a score for a given category can represent the probability that the input image contains an image of an object belonging to that category. As another example, if the input to the verbose machine learning model 110 is a text fragment, the category can be a topic, and a score for a given topic can represent the probability that the input text fragment relates to that topic.

[0020] In some cases, the verbose machine learning model 110 is a single machine learning model. In other cases, the verbose machine learning model 110 is an ensemble machine learning model, which is an assembly of multiple individually trained machine learning models, wherein the outputs of the individual machine learning models are combined to generate the output of the verbose machine learning model 110. Further, in some cases, the models in the ensemble machine learning model include one or more full models that generate scores for each category and one or more specialized models that generate scores only for the corresponding subset of categories. See below for reference. Figure 3 and 4 More specifically, an ensemble machine learning model includes one or more full models and one or more specialized models.

[0021] The model training system 100 trains a refined machine learning model 120 on a set of training inputs to determine the trained parameter values ​​of the refined machine learning model 120 such that the score generated by the refined machine learning model 120 for a given input for a given classification also represents the probability that the classification is the exact classification of the input.

[0022] Specifically, in order to train the refined machine learning model 120, the model training system 100 simultaneously configures the refined machine learning model 120 and the redundant machine learning model 110 to generate soft outputs from the training inputs during the training of the refined machine learning model 120.

[0023] The soft output of a machine learning model for a given input includes a soft score for each category generated by the final layer of the machine learning model. This soft score defines a softer distribution of scores across the set of categories than the scores generated by the machine learning model for that input after the model has been trained.

[0024] Specifically, in some implementations, the final layer of both the refined machine learning model 120 and the verbose machine learning model 110 is a softmax layer, which generates a score q for a given class i. i It satisfies the following equation:

[0025]

[0026] Where z i This is a weighted combination of the outputs of the previous layer of the machine learning model for category i, received by the last layer, where j ranges from 1 to the total number of categories, and T is a temperature constant. In these implementations, the model training system 100 configures both the refined machine learning model 120 and the redundant machine learning model 110 to generate soft outputs by setting T to a higher value than the T used to generate scores after the machine learning model has been trained. For example, after training, the value of T used for the refined machine learning model 120 may be set to equal 1, while during training, the value of T may be set to equal 20. The redundant machine learning model 110 and the refined machine learning model 120 use the same value of T to generate soft outputs during the training of the refined machine learning model 120.

[0027] Therefore, the soft output of a machine learning model is the output of the model using the current values ​​of the parameters, except that the value of T in the last layer of the model is increased to a higher value than the value used to generate the output after the model has been trained, for example, increased to a value greater than 1.

[0028] During training, model training system 100 processes each training input, such as training input 102, using a verbose machine learning model 110 to generate a verbose target soft output for that training input, such as a verbose target soft output 112 for training input 102. Model training system 100 also processes the training input using a refined machine learning model 120 to generate a refined soft output for that training input, such as a refined soft output 132 for training input 102. Model training system 100 then trains the refined machine learning model 120 to generate soft outputs that match the verbose target soft outputs for the training inputs by adjusting the parameter values ​​of the refined machine learning model 120. See below for further details. Figure 2 A more detailed description of training and refining the machine learning model 120.

[0029] Figure 2 This is a flowchart of an example process 200 for training a refined machine learning model. For convenience, process 200 will be described as being performed by a system of one or more computers located in one or more locations. For example, a model training system, such as... Figure 1 The model training system 100, after being properly programmed, can execute process 200.

[0030] This system uses traditional machine learning training techniques to train cumbersome machine learning models, for example, Figure 1 The complex machine learning model 110 is used to determine the trained parameter values ​​of the complex machine learning model (step 202). Specifically, the system trains the complex machine learning model on a set of training inputs to process the inputs and efficiently generate corresponding scores for each category in a predetermined set of categories.

[0031] The system receives training data to train a refined machine learning model, for example, Figure 1 A refined machine learning model 120 (step 204). The training data includes a set of training inputs, which may be the same as or different from the training inputs used to train the complex machine learning model.

[0032] The system uses a trained, verbose machine learning model to process each training input in the training input set to determine a corresponding verbose target soft output for that training input (step 206). As described above, the soft output includes a corresponding soft score for each category. The soft score for a given training input defines a softer score distribution for that input across the category set than the score generated by the machine learning model for that input after the machine learning model has been trained.

[0033] The system trains a refined machine learning model to generate a refined soft output for the training input that matches the verbose target soft output for the training input (step 208).

[0034] Specifically, for a given training input, the system processes the training input using a refined machine learning model to generate a refined soft output for that training input based on the current parameter values ​​of the refined machine learning model. The system then determines the error between the verbose target soft output for that training input and the refined soft output for that training input. The system then uses this error to adjust the parameter values ​​of the refined machine learning model, for example, using traditional machine learning training techniques. For instance, if the refined machine learning model is a deep neural network, the system can adjust the parameter values ​​of the refined machine learning model using gradient descent with backpropagation.

[0035] Optionally, the system can also train the refined machine learning model using a hard target for the training input. The hard target for the training input is a set of scores comprising 1s for each correct or known classification of the training input—that is, each classification the training input should be classified into by the refined machine learning model—and 0s for each other classification. Specifically, the system can determine an error between the verbose target soft output and the refined soft output for the training input, and another error between the hard target and the refined unsoftened output for the training input. The unsoftened output is the set of scores generated by the refined machine learning model for the training input using the temperature that will be used after the model is trained. The system can then use conventional machine learning training techniques to adjust the parameter values ​​of the refined machine learning model using the two errors, for example, by backpropagation gradients computed through the layers of the refined machine learning model using the two errors.

[0036] Figure 3 An example machine learning model system 300 is shown. This machine learning model system 300 is an example of a system implemented as a computer program on one or more computers at one or more locations, in which the systems, components and techniques described below are implemented.

[0037] The machine learning model system 300 uses an ensemble machine learning model 302 to process the input, for example, input 304, to generate a corresponding score for each category in a predetermined set of categories 312.

[0038] The ensemble machine learning model 302 is an ensemble machine learning model that includes one or more specialized machine learning models, such as specialized machine learning model 320 and specialized machine learning model 330, and one or more full machine learning models, such as full machine learning model 310. Each machine learning model in the ensemble machine learning model 302 has been trained separately, for example, using conventional machine learning training techniques, to generate scores representing probabilities.

[0039] Each full machine learning model is configured to process input, such as input 304, and generate a corresponding score for each category in a predetermined set of 312 categories. Although Figure 3 The example only illustrates a single full machine learning model 310; however, the ensemble machine learning model 302 may include multiple full machine learning models, each of which processes the input 304 to generate a corresponding set of scores for that input, each set including the corresponding scores for each category.

[0040] However, each specialized machine learning model was configured to generate scores only for a subset of categories within the 312-category set. Figure 3In the example, the specialist machine learning model 320 is configured to generate scores for a subset 322 of the category set 312, while the specialist machine learning model 330 is configured to generate scores for a subset 332 of the category set 312. In some implementations, each specialist machine learning model may also be configured to generate scores for garbage classification, i.e., a set of all categories not included in the subset of the specialist machine learning model.

[0041] To generate the final score for the classification set of input 304, the machine learning model system 300 processes the input using each full machine learning model to determine an initial set of scores for input 304. The machine learning model system 300 then determines, based on the initial scores for input 304, whether to include scores generated by any specialized machine learning models in the final score for that input 302. See below for reference. Figure 4 The process of processing the input is described in more detail to generate a final set of scores for that input.

[0042] In some cases, the machine learning model system 300 selects subsets to be assigned to each specialized model by selecting subsets such that categories frequently predicted as together by the full machine learning model are included in the same subset. In other words, the machine learning model system 300 can select subsets such that categories frequently recorded as similar by the full machine learning model are assigned to the same subset.

[0043] Specifically, the machine learning model 300 can generate a covariance matrix of scores for the input set processed by the full model, i.e., as described below (refer to...). Figure 4 The initial scores are described. The machine learning model system 300 can then apply clustering techniques, such as K-means clustering, to the columns of the covariance matrix to cluster the categories, with each cluster assigned to a subset of one or more specialties. The machine learning model system 300 can then train a machine learning model for each specialty to efficiently generate scores for that subset of categories assigned to that specialty, and optionally for waste classification.

[0044] Figure 4 This is a flowchart of an example process 400 that processes input using an ensemble machine learning model comprising one or more full machine learning models and one or more specialized machine learning models. For convenience, process 400 will be described as being executed by a system of one or more computers located in one or more locations. For example, a machine learning model system, such as... Figure 3 The machine learning model system 300, after being properly programmed, can execute the process 400.

[0045] The system processes the input using each full machine learning model to generate a corresponding set of full scores for each full machine learning model (step 402). Each set of full scores includes the corresponding full scores for each category in the classification set.

[0046] The system combines the full scores to generate an initial score for the input (step 404). For example, the system can combine the full scores by taking the arithmetic or geometric mean of each full score generated for each category across the entire model. If there is no specialist model in the ensemble machine learning model, the system uses the initial score as the final score for the input.

[0047] If a specialized model is present in the ensemble machine learning model, the system selects a category with a sufficiently high initial score (step 406). The system may determine that the initial score is sufficiently high for each category that has a score exceeding a threshold score, or for each category that has the highest score among the threshold scores.

[0048] The system processes the input using each specialized machine learning model assigned a subset of categories including at least one of the selected categories to generate a corresponding set of specialized scores for the input (step 408). In some embodiments, the system processes the input using each specialized machine learning model and then uses only the scores generated by the specialized machine learning models assigned a subset of categories including at least one of the selected categories. In some other embodiments, after a category has been selected, the system processes the input using only the specialized machine learning models assigned a subset of categories including at least one of the selected categories.

[0049] The system combines the initial score and the specialty score to generate a final score for the input (step 410). For example, the system can combine the score by taking the arithmetic or geometric mean of the initial score generated for that category and the specialty score generated for that category for each category. As another example, the system can combine the score by taking the arithmetic or geometric mean of the full score generated for that category and the specialty score generated for that category for each category.

[0050] When the ensemble machine learning model is as described in the reference above Figure 1 and 2 When training a cumbersome machine learning model for a refined machine learning model, the system sets the temperature to the same higher value for each model in the ensemble machine learning model and then executes process 400 to generate a cumbersome target soft output for a given training input. In some embodiments, the system executes process 400 for the training input to generate an unsoftened cumbersome output for that training input and then generates a cumbersome target soft output by adjusting the unsoftened cumbersome output based on a higher temperature value.

[0051] The above description describes an implementation where the temperature value is constant. However, in some other implementations, the temperature value can change during the process of training the refined machine learning model. For example, the temperature value can gradually increase during the training process, such as after each training iteration. As another example, the temperature value can gradually decrease during the training process, such as after each training iteration.

[0052] The embodiments and functional operations of the subject matter described in this specification can be implemented in digital electronic circuits, tangibly embodied computer software or firmware, computer hardware, including the structures disclosed in this specification and their structural equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by or control of the operation of a data processing apparatus. Alternatively or additionally, the program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are generated to encode information for transmission to a suitable receiver device for execution by the data processing apparatus. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access storage device, or combinations thereof.

[0053] The term "data processing apparatus" includes various means, devices, and machines for processing data, including, for example, programmable processors, computers, or multiple processors or computers. The apparatus may include special-purpose logic circuitry, such as FPGAs (Field-Programmable Gate Arrays) or ASICs (Application-Specific Integrated Circuits). In addition to hardware, the apparatus may also include code that creates an execution environment for the computer program in question, such as code that builds processor firmware, protocol stacks, database management systems, operating systems, or combinations thereof.

[0054] A computer program (also referred to or described as a program, software, software application, module, software module, script, or code) can be written in any form of programming language, including compiled or interpreted languages ​​or declarative or procedural languages, and can be deployed in any form, including as a standalone program, or as a module, component, subroutine, or other unit suitable for use in a computing environment. A computer program may, but is not required to, correspond to a file in a file system. A program may be stored as part of a file that holds other programs or data, such as in one or more scripts in a markup language document, in a single file dedicated to the program in question, or in multiple coordinate files, such as files storing one or more modules, subroutines, or code sections. A computer program can be deployed to execute on a single computer or on multiple computers located at one address or distributed across multiple addresses and interconnected by a communication network.

[0055] The process and logic described in this specification can be executed by one or more programmable computers that execute one or more computer programs to perform functions by manipulating input data and generating outputs. The process and logic can also be executed by dedicated logic circuitry, and the device can be implemented as such dedicated logic circuitry, such as an FPGA (Field-Programmable Gate Array) or an ASIC (Application-Specific Integrated Circuit).

[0056] Computers suitable for executing computer programs include, for example, those based on general-purpose or special-purpose microprocessors or both, or any other type of central processing unit. Generally, the central processing unit receives instructions and data from read-only memory or random access memory or both. The basic components of a computer are the central processing unit for completing or executing instructions and one or more storage devices for storing instructions and data. Typically, a computer will also include, or be operatively coupled to, one or more mass storage devices for storing data, such as magneto-optical, magneto-optical, or optical discs, for receiving data or sending data to said mass storage devices or both. However, a computer is not required to have such devices. Furthermore, a computer may be embedded in another device, such as a mobile phone, personal digital assistant (PDA), mobile audio or video player, game console, global positioning system (GPS) receiver, or portable storage device such as a universal serial bus (USB) flash drive, to name just a few.

[0057] Computer-readable media suitable for storing computer program instructions and data include all forms of non-volatile memory, media, and storage devices, including, for example, example semiconductor storage devices such as EPROM, EEPROM, and flash memory devices; magnetic disks such as internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks. Processors and memory may be complemented or integrated with dedicated logic circuitry.

[0058] To provide interaction with the user, embodiments of the subject matter described in this specification can be implemented on a computer with a display device, such as a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, to display information to the user, and a keyboard and pointing device, such as a mouse or trackball, whereby the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback, such as visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including auditory, voice, or tactile input. Furthermore, the computer can interact with the user by sending documents to and receiving documents from the device used by the user; for example, by sending a webpage to a web browser in response to a request received from a web browser on the user's client device.

[0059] Embodiments of the subject matter described in this specification can be implemented in a computing system that includes backend components, such as a data server, or middleware components, such as an application server, or frontend components, such as a client computer with a graphical user interface or web browser through which a user can interact with embodiments of the subject matter described in this specification, or any combination of one or more such backend, middleware, or frontend components. The components of this system can be interconnected via any form or medium of digital data communication, such as a communication network. Examples of communication networks include local area networks (“LANs”) and wide area networks (“WANs”), such as the Internet.

[0060] The computing system may include clients and servers. Clients and servers are typically geographically separated and interact via a communication network. The relationship between clients and servers is manifested by computer programs running on their respective computers and having a client-server relationship with each other.

[0061] While this specification includes numerous details of specific implementations, these should not be construed as limiting the scope of any claimable invention, but rather as descriptions of features that may be specific to particular embodiments of a particular invention. Certain features described in the context of individual embodiments of this specification may also be implemented together in a single embodiment. Conversely, multiple features described in the context of a single embodiment may also be implemented individually in multiple embodiments or in any suitable sub-combination. Furthermore, while the features described above may be described as being used in certain combinations, and even in the initial declaration, one or more features in a declared combination may be detached from that combination in certain circumstances, and a declared combination may refer to a sub-combination or a variation of a sub-combination.

[0062] Similarly, although the operations are described in a specific order in the figures, this should not be construed as requiring these operations to be executed in the specific or sequential order shown, or requiring all illustrated operations to be executed to obtain the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of multiple system modules and components in the embodiments described above should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0063] Specific embodiments of this subject matter have been described. Other embodiments are also within the scope of the claims. For example, the actions referenced in the claims can be performed in a different order and still obtain the desired result. As an example, the processes described in the drawings are not necessarily required to follow the specific order or sequential order shown to obtain the desired result. In some embodiments, multitasking and parallel processing may be advantageous.

Claims

1. A system for training an image processing model, the system comprising one or more computers and one or more storage devices storing instructions, the instructions, when executed by the one or more computers, causing the one or more computers to perform a method, the method comprising: Training a complex machine learning model, wherein the complex machine learning model is configured to receive input image data representing an image and generate a corresponding score for each of a plurality of categories, wherein the score represents the probability that the image contains an object belonging to that category; and A refined machine learning model is trained on multiple training images, wherein the refined machine learning model is also configured to receive input image data representing images and generate a corresponding score for each of the multiple categories, wherein the score represents the probability that the image contains an object belonging to that category, including: Each training image is processed using the complex machine learning model to generate a complex target soft output for that training image; and The refined machine learning model is trained to generate a soft output for each of the plurality of training images that matches the verbose target soft output for that training image. Wherein, the verbose target soft output and the soft output matching the verbose target soft output each include a corresponding soft score for each of the plurality of classifications generated by the last layer of the respective machine learning model, and wherein each soft score satisfies: , where q i is a score for a classification i, z i is a weighted combination of outputs of a preceding layer of the respective machine learning model received by the last layer, j ranges from 1 to a total number of classifications in the plurality of classifications, T is a temperature constant, and wherein, to generate the soft score, T is set to a value higher than a value used to generate scores for the classifications after the respective machine learning model has been trained.

2. The system of claim 1, wherein the value of T used to generate a score for the classification after the machine learning model has been trained is 1.

3. The system of claim 2, further comprising setting the value of T to 1 for the refined machine learning model after the refined machine learning model has been trained.

4. The system of claim 1, wherein the soft output of the machine learning model comprises a corresponding soft score for each of the plurality of categories generated by the last layer of the machine learning model, and wherein the soft score defines a softer score distribution across the plurality of categories compared to scores generated by the machine learning model after the machine learning model has been trained.

5. The system of claim 1, wherein the cumbersome machine learning model is an ensemble model composed of multiple machine learning models, and wherein the corresponding score generated by the cumbersome machine learning model for each category is a combination of the scores generated by the machine learning model for that category.

6. The system of claim 5, wherein the plurality of ensemble models are trained separately.

7. The system of claim 5, wherein processing each training image using the verbose machine learning model comprises: The ensemble model is used to generate multiple ensemble scores for the training images; as well as The integrated scores are combined to generate the verbose target soft output for the training images.

8. The system of claim 7, wherein the plurality of machine learning models comprises one or more full machine learning models and one or more specialized machine learning models, wherein each full machine learning model is configured to receive input and generate a corresponding full score for each of the plurality of categories, and wherein each specialized machine learning model is configured to receive input and generate a corresponding specialized score for a corresponding subset of the plurality of categories.

9. The system of claim 8, wherein each specialized machine learning model is further configured to generate a specialized score for waste sorting, said waste sorting being an aggregation of all sorts not included in said subset used for said specialized machine learning model.

10. The system of claim 8, wherein combining the ensemble scores to generate the verbose target soft output for the training images comprises: The full scores generated by each of the full machine learning models for the training image are combined to generate an initial score for the training image; as well as For each category with a sufficiently high initial score, the full score for that category and the professional score generated by the professional model for that category are combined, wherein the category is included in the subset of categories used for the professional model.

11. The system of claim 8, further comprising: One or more subsets of the multiple categories are generated by clustering the categories frequently predicted together by the one or more full machine learning models into the same subset; as well as Each subset is assigned to a corresponding specialized model.

12. The system of claim 1, wherein training the refined machine learning model on the plurality of training images further comprises: The refined machine learning model is trained to generate unsoftened outputs that match the known outputs for each of the plurality of training images.

13. A system for training an image processing model, the system comprising one or more computers and one or more storage devices storing instructions, the instructions, when executed by the one or more computers, causing the one or more computers to perform a method, the method comprising: Train a refined machine learning model with multiple refined model parameters on the training input set, where: During the training period, the refined machine learning model is configured to process the input by performing operations including the following to generate an output score distribution across multiple categories: Process the input to generate an intermediate score distribution across the multiple categories; and The intermediate score distribution is processed using the final layer of the refined machine learning model to generate the output score distribution. The final layer is parameterized by a temperature parameter with training values, wherein the temperature parameter controls the softening degree of the output score distribution; and The training includes: Each training input in the training input set is processed using a complex machine learning model to generate a target score distribution across the multiple categories; and The refined machine learning model is trained for each training input in the training input set, and the training input is processed to generate an output score distribution that matches the target score distribution generated by the verbose machine learning model for the training input; Wherein, after the training, the temperature parameter of the final layer of the refined machine learning model is configured to have an inference value that is different from the training value of the temperature parameter used during training.

14. A method for training an image processing model, the method comprising operations performed by the corresponding system according to any one of claims 1-13.

15. A computer storage medium encoded with instructions that, when executed by one or more computers, cause the one or more computers to perform the operation of the method according to claim 14.

Citation Information

Patent Citations

  • System and Method for Combining Segmentation Data

    US20140122401A1