CN111556998A - Transfer learning and domain adaptation using distributable data models - Google Patents

Abstract

A system for using a distributable data model for transport learning and domain adaptation is provided, comprising, a networked distributable model configured to serve a plurality of distributable model instances; and a directed computational graph module configured to receive at least an instance of the at least one distributable model from the networked computing system, create a second dataset from machine learning performed by the transport engine, train the instance of the distributable model with the second dataset, and generate an update report based at least in part on an update to the instance of the distributable model.

Description

Transfer Learning and Domain Adaptation Using Distributable Data Models

Cross-references related to the application

This application is a PCT application of US Patent Application Serial No. 15/835,436 filed on December 7, 2017, entitled "TRANSFER LEARNING AND DOMAINADAPTATION USING DISTRIBUTABLE DATA MODELS", and claims priority thereof, which is hereby incorporated by reference in its entirety. Its entire description is incorporated herein.

technical field

The present disclosure relates to the field of machine learning, and more particularly to model improvement using biases contained in data distributed across multiple devices.

Background technique

In traditional machine learning, data is often collected and processed in a central location. The collected data can then be used to train the model. However, the data that can be collected through means such as web scraping or news aggregation is relatively narrow in scope compared to crime data stored on devices, such as personal mobile devices, or from local police departments. This data can be difficult to leave the device in which it is stored due to the likelihood that it contains sensitive data, and the bandwidth required to transmit the data can also be substantial.

Generator models (whether anonymous or not) trained on the data line can be restricted for transmission (eg due to GDPR). Existing tools such as SNORKEL ^™ can provide substantial mechanisms to accelerate the generation of realistic and labeled training data from small seeds. The same concept can support the transfer of valuable models without moving the restricted information itself. This also applies to the sharing of other models, datasets, visualizations, pipelines, and other data, such as multi-tenant deployments, inter-organization sharing, or hybrid cloud edge use cases.

What is needed is a system for using transfer learning to fit a data model trained in a particular environment or application that may not have an equivalent feature space or distribution within the target data or application.

SUMMARY OF THE INVENTION

Accordingly, the inventors have conceived a system and method for using transfer learning to fit a data model trained in a particular environment or application that may not have an equivalent feature space or distribution within the target data or application.

In a preferred embodiment, the model source may serve instances of various distributable models: generalized models in which the data used to train them have biases weighted and corrected, and bias-specific models in which biases are exploited. Instances serve distributed devices, where they can be trained by devices. Each device generates an update report, which is transmitted back to the model source, where it can be used by the model source to improve the main model.

In one aspect of the invention, there is provided a system for improving a distributable model having biases contained in distributed data, comprising a networked distributable model source comprising a memory, a processor, and storage in its memory a plurality of programming instructions executable on its processor, wherein the programmable instructions, when executed on the processor, cause the processor to serve a plurality of instances of the distributable model; and a directed computational graph module including memory, processing a processor and a plurality of programming instructions stored in its memory and executable on its processor, wherein the programmable instructions, when executed on the processor, cause the processor to receive at least an instance of the distributable model from a networked computing system, creating a sanitized dataset from the data stored in the memory based at least in part on biases contained in the data stored in the memory, training an instance of the distributable model with the sanitized dataset, and at least in part by updating the distributable An instance of the model is used to generate an update report.

In another embodiment of the invention, at least a portion of the sanitized data set is data from which sensitive information has been removed. In another embodiment of the invention, at least a portion of the update report is used by a networked distributable model source to improve the distributable model. In another embodiment of the invention, at least a portion of the deviation contained within the data stored in the memory is based on geographic location. In another embodiment of the invention, at least a portion of the deviation contained in the data stored in the memory is based on age. In another embodiment of the invention, at least a portion of the variance contained in the data stored in the memory is based on gender.

In another characteristic aspect of the invention, there is provided a method for improving a distributable model using biases contained in distributed data, comprising the steps of: (a) serving a plurality of distributable models using a networked distributable model source an instance of a distributed model; (b) receiving at least one instance of at least one distributable model from a networked computing system having a directed computational graph module; (c) based at least in part on storage contained in a memory having a directed computational graph module (d) using a directed computational graph module to train distributable model instances of the sanitized dataset; and (e) at least in part by employing The directed computational graph module updates the distributed model instance to generate an update report.

Description of drawings

The drawings illustrate several characteristic aspects, and together with the description serve to explain the principles of the invention in accordance with the characteristic aspects. It should be appreciated by those skilled in the art that the specific arrangements shown in the figures are exemplary only and should not be construed as limiting the scope of the invention or the claims herein in any way.

FIG. 1 is an exemplary architecture diagram of a service operating system according to an embodiment of the present invention.

2 is a block diagram of an exemplary system for using transfer learning to fit a data model trained in a particular environment or application that may not have an equivalent feature space or distribution within the target data or application, according to various embodiments of the present invention.

3 is a block diagram of an exemplary system that can exploit bias contained within data distributed among sectors to improve sector-level distributable models, in accordance with various embodiments of the present invention.

4 is a block diagram of an exemplary hierarchy in which each hierarchy may have its own distributable model, according to various embodiments of the present invention.

5 is a flowchart illustrating a method for cleaning and decrypting data stored on an electronic device prior to training a distributable model instance in accordance with various embodiments of the present invention.

6 is a flowchart illustrating a method for improving a distributable model on a model source external device in accordance with various embodiments of the present invention.

7 is a flow diagram illustrating a method for improving generic and bias-specific distributable models using biases obtained from data on distributed devices, in accordance with various embodiments of the present invention.

8 is a flowchart illustrating a method for using transfer learning to fit a data model trained in a particular environment or application that may not have an equivalent feature space or distribution within the target data or application.

9 is a flowchart illustrating a method for using a pretrained model in transfer learning to provide a starting point for additional model training.

10 is a block diagram illustrating an exemplary hardware architecture of a computing device for various embodiments of the present invention.

11 is a block diagram illustrating an exemplary logical architecture for a client device in accordance with various embodiments of the present invention.

12 is a block diagram illustrating an exemplary architectural arrangement of clients, servers, and external devices in accordance with various embodiments of the present invention.

13 is another block diagram illustrating an exemplary hardware architecture of a computing device for various embodiments of the present invention.

Detailed ways

The present inventors have envisioned and put into practice a system and method for improving regionalized distributable models using biases contained in distributed data.

One or more different feature aspects may be described in this application. Further, many alternative arrangements may be described for one or more of the feature aspects described herein; it should be understood that these arrangements are presented for illustrative purposes only and are not in any way limiting of the feature aspects contained herein or herein Presented claims. One or more arrangements are broadly applicable to many features, as can be readily seen from this disclosure. Typically, arrangements are described in sufficient detail to enable those skilled in the art to practice one or more of the feature aspects, and it is to be understood that other arrangements may be used and made without departing from the scope of the particular feature aspects Structural, logical, software, electrical and other changes. Specific features of one or more feature aspects described herein may be described with reference to one or more specific feature aspects or figures that form part of this disclosure and that diagrammatically illustrate specific arrangements of one or more aspects. It should be understood, however, that these features are not limited to their use in the context of one or more of the specific features or figures in which they are described. The invention is neither a literal description of all arrangements of one or more feature aspects nor a feature listing of one or more feature aspects that must be present in all arrangements.

The paragraph headings and the headings of this patent application are provided in this patent application for convenience only and should not be construed as limiting the present disclosure in any way.

Devices that are in communication with each other need not be in continuous communication with each other unless expressly specified otherwise. Furthermore, devices that are in communication with each other may communicate logically or physically, either directly or indirectly through one or more communication mechanisms or intermediaries.

A description of an aspect having multiple components in communication with each other does not imply that all of these components are required. Rather, various optional components may be described to illustrate various possible feature aspects, and in order to more fully describe one or more feature aspects. Similarly, although process steps, method steps, algorithms, etc. may be described in sequence, unless expressly stated to the contrary, such processes, methods, and algorithms may generally be configured to operate in alternate sequences. In other words, any sequence or order of steps described in this patent application does not by itself imply that the steps are required to be performed in that order. The steps of the described process may be performed in any practical order. Further, although described or suggested as non-concurrent, some steps may be performed concurrently (eg, because one step is described after another step). Furthermore, the depiction of a process in the accompanying figures does not imply that the process shown excludes other variations or modifications thereof, nor that the process shown or any step thereof is necessary for one or more features, nor There is no suggestion that the process shown is preferred. Furthermore, steps are typically described once per feature, but this does not imply that they must occur once, or that they may occur only once per execution or execution of a process, method, or algorithm. Certain steps may be omitted in certain aspects or events, or certain steps may be performed multiple times in a given aspect or event.

When a single device or item is described herein, it will be apparent that more than one device or item may be used in place of a single device or item. Similarly, when more than one device or item is described herein, it will be apparent that a single device or item may be used in place of more than one device or item.

A function or feature of a device may alternatively be embodied by one or more other devices not expressly described as having that function or feature. Therefore, other characteristic aspects need not necessarily include the device itself.

Techniques and mechanisms described or referenced herein are sometimes described in the singular for clarity. It should be appreciated, however, that a particular feature aspect may include multiple iterations of the technique or multiple instances of the mechanism, unless otherwise stated. Process descriptions or blocks in the figures should be understood to represent modules, code segments, or portions of code that comprise one or more executable instructions for implementing the particular logical functions or arrangements in the process. Alternative implementations are included within the scope of the various feature aspects, in which, for example, the functions may be performed out of the order shown or discussed, including substantially concurrently or in the reverse order, depending on the functionality involved, as skilled in the art People should understand.

conceptual architecture

FIG. 1 is an exemplary architecture diagram of a service operating system 100 according to an embodiment of the present invention. Client access to the system 105 for ad hoc data input, system control, and for interaction with system outputs such as automated predictive decisions and planning and alternative path simulations occurs through the system's distributed, scalable high-bandwidth cloud interface 110 and data storage 112 , interface 110 uses a multipurpose, robust web application driven interface to input and display client-facing information, data store 112 such as but not limited to MONGODB ^™ , COUCHDB ^™ , CASSANDRA ^™ or REDIS ^™ , depending on the embodiment. Most of the business data analyzed by the system comes from both data sources within the customer's business as well as from cloud-based, public or proprietary data sources107 such as, but not limited to: subscribed business domain-specific data services, external remote sensors , subscribed satellite imagery and data feeds and websites of general and domain-specific commercial operations of interest, also enter the system through the cloud interface 110, data is passed to the API routines 135a that may have the necessary to receive and transform external data, and then The connector module 135 , the directed computational graph module 155 , the high-volume web crawler module 115 , the multidimensional time series database 120 and the graph stack service 145 that pass the normalized information to other analysis and transformation components of the system. The directed computational graph module 155 retrieves one or more data streams from multiple sources including, but not limited to, multiple physical sensors, network service providers, web-based questionnaires and surveys, monitoring of electronic infrastructure, mass events and human input device information. Within the directed computational graph module 155, the data can be divided into two equivalent streams in a dedicated preprogrammed data pipeline 155a, where one substream can be sent for batch processing and storage, while the other substream can be reformatted for transform pipeline analysis. The data may then be passed to a general transformer service module 160 for linear data transformations as part of the analysis, or a decomposable transformer service module 150 for branching or iterative transformations as part of the analysis. The directed computational graph module 155 represents all data as a directed graph, where the transforms are nodes, and the resulting messages are represented between the transform edges of the graph. The high volume web scraping module 115 may use a number of pre-programmed web crawlers residing on the server, which crawlers, although self-configured, may be deployed within the web scraping framework 115a, of which SCRAPY ^™ is an example, for identifying and retrieving data of interest from web-based data sources that are not well-labeled by traditional web crawling techniques. The multidimensional time series data storage module 120 may receive streaming data from a large number of sensors, which may be of several different types. The multidimensional time series data storage module 120 may also store any time series data encountered by the system 100, such as, but not limited to, environmental factors at the insured client infrastructure site, component sensor readings and system logs for some or all of the insured clients, the insured client Weather and catastrophic event reports in the region where the client is located, political bulletins and/or news from regions where information on insured customer infrastructure and network services resides (such as, but not limited to, news, capital financing opportunities and financial information, sales , market conditions), and customer data related to the Services. The multi-dimensional time series data storage module 120 can accommodate irregular and large-capacity surges by dynamically allocating network bandwidth and server processing channels to process incoming data. Examples of programming wrappers 120a for languages include, but are not limited to, C++, PERL, PYTHON, and ERLANG ^™ , which allow complex programming logic to be added to the default functionality of multidimensional time series database 120 without being familiar with core programming, extremely Greatly expands the breadth of functionality. The data retrieved by the multi-dimensional time series database 120 and the bulk web crawler module 115 can be further analyzed and transformed into task optimization results by the directed computation graph 155 and associated generic transformer service 160 and decomposable transformer service 150 modules. Alternatively, data from multidimensional time-series databases and bulk web crawler modules (usually with script hints 145a identifying important vertices) can be sent to the graphics stack service module 145, which converts the flow of information into this data using a standardized protocol A graphical representation of , such as Open Graph Internet Technology (although the present invention does not rely on any one standard). Through these steps, the graph stack services module 145 graphically represents the data affected by any predetermined script modification 145a and stores it in a graph-based data store 145b, such as GIRAPH ^™ or key-value paired data store REDIS ^™ , or RIAK ^TM , either of which is suitable for storing graph-based information.

The results of the transformation analysis process can then be combined with further client-side instructions, additional business rules and practices related to the analysis, and situational information beyond the existing data in the automated planning services module 130, which also runs powerful information-based Theoretical predictive statistics functions and machine learning algorithms 130a to allow rapid prediction of future trends and outcomes based on results derived from current systems and selection of each of a number of possible business decisions. Then, using all or most of the available data, the automatic planning service module 130 can make business decisions that are most likely to lead to favorable business outcomes with a high level of certainty that can be used. Closely related to the automation planning service module 130, the action result simulation module 125 with the discrete event simulator programming module 125a is associated with end-user-oriented observations and A state estimation service 140 is coupled, which has a highly scriptable 140b as the situation demands, and a game engine 140a to more realistically implement the likely outcomes of the business decision under consideration, allowing business decision makers based on analysis of currently available data And investigate the possible outcomes of choosing one pending course of action over another.

2 is a block diagram of an exemplary system 200 for using transfer learning to fit a data model trained in a particular environment or application that may not have equivalent feature spaces or portions within the target data or application, according to various embodiments of the present invention . System 200 includes a distributable model source 201, and a plurality of electronic devices 220[a-n]. Model source 201 may include distributable models 205, local data store 210, and synthetic data store 215; while each electronic device may have instance models 221[a-n], and device data 222[a-n]. Electronic devices 220[a-n] may be any type of computerized hardware commonly used in the art, including, but not limited to, desktop computers, laptop computers, mobile devices, tablet computers, and computer clusters. For simplicity, the following discussion of system 200 will be from the perspective of a single device 220a, which includes instance model 221a and device data 222a. It should be understood that devices 220[a-n] may operate independently and in a similar manner to device 220a.

The model source 201 may be configured to send an instance copy of the distributable model 205 to the device 220a, and may be any type of computerized hardware used in the art configured for use with the business operating system 100, and may be processed over the air via a A directed computational graph data pipeline 155a for incoming and outgoing data communicates with connected devices. The distributable model 205 may be a model commonly used in machine learning in the art, or it may be a specialized model that has been tuned to more efficiently utilize training data that has been distributed among devices and is still accessible through traditional machine learning method to train. Local data 210 may include data that was previously stored, or data that is in the process of being stored, eg, data currently being collected by monitoring other systems. Synthetic data 215 may be data that has been intelligently and/or predictively generated to fit the model context, and is typically based on real trends and events. Examples of synthetic data 215 may include data that has been collected from a running computer simulation; data that has been generated using predictive simulation capabilities of the business operating system 100 with previously stored data and current trends; or employing other specialized software, such as SNORKEL. Local data 210 and synthetic data 215 may be used by model source 201 as training data for distributable model 205 . Although shown in Figure 2 as being stored within model source 201, those skilled in the art will appreciate that local data 210 and synthetic data 215 may originate from external systems and be provided to model source 201 through some mechanism such as, for example, an Internet or local area network connection .

In system 200, device 220a may be connected to model source 201 via a network connection such as the Internet, a local area network, a virtual private network, etc., where device 220a may provide an instance copy of distributable model 221a. Device 220a may then sanitize and clean its own data 222a, and train its instance model 221a using the sanitized data. Reports based on updates to the instance model 221a may be used by the model source 201 to improve the distributable model 205 . In a preferred embodiment, the apparatuses 220[a-n] may comprise a system configured to use the business operating system 100 and utilize its directed computational graph capabilities among other functions to process data. In some embodiments, however, the hardware of the electronic device may be, for example, a mobile device or a computer system running another operating system. These systems may use specialized software configured to process data according to established compatibility specifications.

The transport engine 230 may be used to facilitate the sharing of the model 205 with some destination devices 220b if desired, eg, for use with the adaptive model 205 for a different environment from which it was originally created or for a different purpose. This can be used, for example, to share data models between departments or organizations, which can benefit from mutual data regardless of changes in hardware or software environments or different purposes. The transfer engine 230 may utilize various machine learning schemes to provide transfer learning, taking machine learning from a problem or set of data and applying it to different (but related) data. For example, machine learning trained on a set of data about automobiles may be applicable to data about other vehicles such as trucks or motorcycles, and transfer learning may be used to enable the application of automobile-specific data to broader data sets and related data types. Transfer learning schemes may include, for example, probabilistic logic such as Markov logic networks (which can be used to apply learned techniques to new data models, or to modify data models for new purposes, in order to facilitate inference under uncertainty) or Bayesian model to model various conditional dependencies as directed acyclic graphs. These approaches, and others that may be possible in terms of various characteristics, can be used to analyze a data set according to the intended purpose for the data set, such as, for example, ensuring that the data is compliant with any applicable regulations (such as reviewing possible personal availability identifying information, for example) or ensuring that the data conforms to any required specifications to ensure interoperability with the intended destination or use.

The transfer engine 230 can use transfer learning to adapt to environments or applications that may have different characteristics or meet different criteria so that the suitability of the data set for a particular destination can be assessed. Analysis can be further enhanced by using multiple pre-trained models that can be used as a starting point for further refinement during operation, applying machine learning algorithms to the model itself so that it can be modified to fit the data or destination information changes, ensuring that the analysis remains fit for the intended design purpose and the dataset being analyzed. In addition, pre-trained models can be used to adapt to environments that do not require in-depth analysis, such as (for example) in resource-constrained deployments (eg embedded or mobile devices), where they can be exported for storage or used with more advanced The machine learning capability uses pre-trained models to perform initial analysis of the data before. This also provides a mechanism for adapting to environments where localization training may not be required, such as when deploying across many mobile devices, it may be necessary to ensure consistency of analysis operations, and it is not possible for each device to develop its own localization Train the model.

The transfer engine 230 may be used to facilitate domain adaptation, learning from one dataset (the initial distributable data model 205), and applying this machine learning to new data to create a new data model based on the knowledge gained from the initial training. This can be performed in a partially or fully unsupervised manner, learning from source data without manually labeling the data (as is commonly used in traditional approaches). In a partially unsupervised arrangement, a pretrained model with some labeled data can be used as a starting point for further training, where unsupervised learning incorporates the labeled information and builds on it to develop a larger model for use. This can be used to incorporate existing labeled data (for example, data that may have been created or labeled for other applications), and then apply unsupervised machine learning to extend the data without further manual action.

The system 200 may also be configured to allow certain deviations in the data to remain unchanged. In some cases, deviations such as the geographic origin of the data may be necessary to comply with local laws. For example, some countries may have certain laws in place to prevent the export and import of data to other restricted countries. Bias can be used to classify incoming update reports by taking into account the bias of the geographic origin of the data. The updated report can then be used to train a geographically restricted distributable model without including data from other restricted countries, while still including data from non-restricted countries.

To provide another exemplary application, crime reports and data distributed across multiple policing municipalities may contain information on emerging trends in crime that can be correlated to crime prediction models. However, this data can contain sensitive, non-public information that must remain within the urban area. With the system described herein, data can be removed from sensitive information, instances used to train a crime prediction model, and relevant data can be indirectly integrated into the crime prediction model. Because the actual data doesn't leave the city's computer systems, this can be done without fear of leaking sensitive information.

It should be understood that the electronic devices 220[a-n] need not all be connected to the model source 201 at the same time, and that any number of devices may be actively connected to and interact with the model source 201 at a given time. Furthermore, the refinement of the distributable model, the refinement of the instance model, and the back-and-forth communication can run discontinuously, and can be performed only when needed. Furthermore, for simplicity, only one example is shown in FIG. 2 . In practice, however, system 200 may be a segment of a potentially larger system having multiple distributable model sources; or a distributable model source may store and serve multiple distributable models. The electronic devices 220[a-n] are not limited to interacting with only one distributable model source, and may connect and interact with as many distributable model sources as desired.

Because biases contained within the data can contain valuable information or give valuable insight when applied to the correct model, it can be prudent to use this data for specialized models.

FIG. 3 is a block diagram of an exemplary system 300 that can utilize the bias contained in data distributed among various sectors to improve a bias-specific distributable model in accordance with various embodiments of the present invention. System 300 may include a distributable model source 301; and a set of sectors 302, which may include a plurality of sectors 325[a-h]. Model source 301 may be a system similar to model source 201, including a source of local data 310 and a source of synthetic data 315 that may be used in an equivalent manner; however, model source 301 may serve instances of two different models - Generalized Distributable Model 305 and bias-specific distributable model 320. The generalized distributable model 305 may be a model that can be trained using data distributed across various sectors in the group 302 , where within-data biases have been weighted and corrected, similar to the distributable model 205 . The bias-specific distributable model 320 may be one in which any bias contained within the data from each sector 325[a-n] may be carefully considered.

Each of the sectors 325[a-h] may be devices with distributed data, similar to the electronic devices 220[a-n] in Figure 2, which have been assigned a classification, eg, devices within a certain area. However, sectors are not limited to being based on geographic location. To provide some non-geographic examples, sectors 325[a-h] may be devices categorized by age group, income category, interest group, and the like. Devices within any of sectors 325[a-h] may obtain instances of the generalized distributable model and the bias-specific model. The device can then train the instance model using its own stored data, which has been processed in the manner described below in Figure 7, and which ultimately leads to improvements in the generalized model and the bias-specific model.

It should be appreciated that, similar to the system 200 shown in FIG. 2, the embodiment shown in the system 300 shows one possible arrangement of distributable model sources and data sources, and this arrangement has been chosen for simplicity. In other embodiments, there may be, for example, more or fewer sectors in a group, there may be multiple groups of sectors, or there may be multiple models, both generalized and bias-specific. See discussion of Figure 4 below for an example with multiple layers, each layer having its own group of sectors.

4 is a block diagram of an exemplary hierarchy 400 in which each layer may have its own set of distributable models, according to various embodiments of the present invention. At the highest level of hierarchy 400, there is a global hierarchy 405, which may contain all lower levels of hierarchy 400. Below the global level 405 may be a number of continent level 410 sectors, which in turn may include all levels below it. Continuing this pattern, there can be multiple national 415 sectors below the continent level, where there can be multiple national or provincial 420 sectors, which results in county 425 sectors, then city 430 sectors, Then there is the street level 435 sector. Each level of sector may have a system as described in Figure 3, with a distributable model dedicated to exploiting biases collected from data within each level of sector distribution, and with some degree of weighted and corrected biases The generalized distributable model of . For example, a model can also be shared between tiers if a particular model on one tier applies to another tier. Models and data collected within a tier may not be limited to use in adjacent tiers, eg, by weighting and correcting for bias, data collected at the county level may also correlate to models at the global level as well as models at the street level.

For example, a global company may have business offices around the world. The company can have a global office that manages offices in a subgroup of business areas, such as commonly used business areas such as Asia Pacific, US, Latin America, Southeast Asia, etc. These business areas can each have a regional office that manages their own country groupings within those areas. These countries may each have a country office that manages their own state/province/territory groupings. As shown in Figure 4, there can be deeper, finer-grained subgroups, but for the purposes of this example, the deepest level used will be state/province/territory.

A global office may have systems configured to use the business operating system 100 and may serve two distributable models: a generalized distributable model configured to analyze sanitized data from distributed business areas on a global scale, and A bias-specific distribution model that exploits bias in data distributed between business areas. An example of a global model could be a model that predicts impacts on the environment or global climate, while a business-region-specific model could be a model that predicts business benefits based on regional data, such as the economy of each region, the development trends of each region, or the s population. At the next level, regional offices can provide distributable models to countries within the region: distributed models for bias-corrected region-wide distributed data, and bias-based distributable exploiting biases within country-level distributed data Model. These models can be similar to those served by the Global Office, but can be more regionally and country-centric, for example, regional environment and climate, and country-level business forecasts. At the next level, and the final level in this example, the country office can provide a distributable model to the states/provinces/territories within the country, similar to the example models at the global and regional levels. Data distributed at the state level can be collected and used to train these models. It should be noted that the distribution of the models is divided at each level to simplify the example and does not represent any limitation of the invention. In other embodiments, for example, a single system at any level may serve all of the various models at each level of the hierarchy, or there may be servers serving models outside of the global company's system used in this example .

DETAILED DESCRIPTION OF EXEMPLARY FEATURES

5 is a flowchart illustrating a method 500 for sanitizing and deciphering data stored on an electronic device prior to training an instance of a distributable model, according to various embodiments of the present invention. In an initial step 505, biases found within the data are identified and corrected so that the data may be more favorable for use in a generalized distributable model. Bias can include, but is not limited to, ontological or regional trends, linguistic or ethnic trends, and gender trends. Bias is intelligently weighted to provide useful insights without overfitting the dataset. In some embodiments, bias data may be specifically identified and stored for use on other, more localized, distributable models. At step 510, in order to maintain user and system privacy, the data may be cleared of personally identifiable information and other sensitive information. This information may include banking information, medical history, addresses, personal history, and the like. At step 515, the now sanitized and scrambled data can be marked for use as training data. At step 520, the data is now formatted appropriately for training the instance model.

It should be noted that method 500 outlines a method of sanitizing and deciphering data. In other embodiments, some steps in method 500 may be skipped, or other steps may be added. In addition to being used to train instance models, and ultimately improve distributable models, data can also be processed in a similar way to become more general and suitable for transfer in training models belonging to different domains. This allows previously collected data to be reused in new models without the sometimes laborious re-collection of data formatted for the context of the newly created model's specific domain.

6 is a flowchart illustrating a method 600 for improving a distributable model from a model source on an external device, according to various embodiments of the present invention. In an initial step 605, the instance of the distributable model is shared with devices from the source of the distributable model. At step 610 , data on the device is sanitized and declassified, one such method is outlined in FIG. 5 . At step 615, the data can now be used to train the shared model (ie, the shared instance of the distributable model) on the device, in a format suitable for training a generic distributable model. At step 620, a report based on the updates to the shared model can be generated and transmitted to the distributable model source. One benefit of generating and transmitting a report, as opposed to transmitting the data itself, is that it may be more efficient than transmitting all the data needed to improve the distributable model. Additionally, since the raw data does not need to leave the device, this approach may help further ensure that sensitive information is not leaked. At step 625, the distributable model source uses the report to improve the distributable model.

7 is a flowchart illustrating a method 700 for improving bias-specific and generalized distributable models using data on distributed devices, according to various embodiments of the present invention. In an initial step 705, the apparatus obtains instances of distributable models from a model source: generalized distributable models and bias-specific distributable models. At step 710, a first dataset is created using data stored on a device that has biases in the data intelligently identified and processed by the business operating system 100 for use with bias-specific models. At step 715, a second dataset is created using the data stored on the device, with biases in the data intelligently weighted and corrected so that the dataset may be more suitable for the generalized model. At step 720, the dataset is sanitized and sensitive information removed. At step 725, the first and second datasets are used to train instances of their respective models; the first dataset trains instances of the bias-specific model, and the second dataset trains instances of the generalized model. At step 730, a report is generated based on the updates made to each instance of the model. After the report is generated, the steps are similar to the final steps of method 600, in particular steps 620 and 625. The generated reports are transferred to the model source, each report is used to improve the respective model.

8 is a flowchart illustrating a method 800 for using transfer learning to fit a data model trained in a specific environment or application that may not have the same feature space or distribution in the target data or application . In an initial step 801, a first dataset may be created from stored data, optionally with deviations identified or removed in accordance with one or more of the features previously described (refer to Figures 4-7). In the next step 802, the destination of the dataset (eg, an external application or service for which the data is intended, or another environment or department with which the data is shared) can be analyzed to identify any requirements for the dataset, such as ( For example) regulatory, content, or format restrictions to which the data should adhere. In the next step 803, various transfer learning algorithms can be applied to the first data set, using machine learning to process the data and determine its suitability for the destination 804, and identify areas where the data must be adapted to meet the requirements. A number of transfer learning methods can be used according to specific arrangements, for example using probabilistic learning models such as Bayesian or Markov networks to appropriately handle data transfers for any uncertainty that may exist. In a next step 805, the results of the transfer learning process can be used to create a second dataset, generating a derived dataset containing appropriate and adaptive data from the first dataset, such that the second dataset is fully compatible with the destination. In a final step 806, the second data set may be provided to the data destination for use.

9 is a flowchart illustrating a method 900 of using a pretrained model in transfer learning to provide a starting point for additional model training. In an initial step 901, a first dataset may be created from stored data, optionally with deviations identified or removed in accordance with one or more of the aspects previously described (refer to Figures 4-7). In the next step 902, the destination of the dataset (eg, an external application or service for which the data is intended, or another environment or department with which the data is shared) can be analyzed to determine any requirements for the dataset, such as ( For example) regulatory, content, or format restrictions to which the data should adhere. In a next step 903, a pretrained model containing labeled data can be loaded, providing an initial model to use as a starting point for unsupervised learning based on a portion of the previously labeled data (as opposed to traditional, fully supervised approaches, where all data must be manually marked for machine learning operations). In the next step 904, unsupervised machine learning can then process the first dataset from the labeled data in the pre-trained model, built from the labeled data to develop a second data model 905 for transfer and domain adaptation.

hardware architecture

Generally, the techniques disclosed herein may be implemented on hardware or a combination of software and hardware. For example, they can be implemented in the operating system kernel, in a separate user process, in a library package bound to a network application, on a specially constructed machine, on an application specific integrated circuit (ASIC), or on a network interface card superior.

Hybrid software/hardware implementations of at least some of the features disclosed herein may be implemented on a programmable network-resident machine (understood as including intermittently connected network-aware machines) selected by a computer program stored in memory be activated or reconfigured automatically. The network device may have multiple network interfaces that may be configured or designed to utilize different types of network communication protocols. The general architecture of some of these machines may be described herein in order to illustrate one or more exemplary means by which a given functional unit may be implemented. According to certain feature aspects, at least some of the features or functions of the various feature aspects disclosed herein may be implemented on one or more general-purpose computers, such as, for example, end-user computer systems, client computers, associated with one or more networks , web server or other server system, mobile computing device (eg, tablet, mobile phone, smartphone, laptop or other suitable computing device), consumer electronic device, music player or any other suitable electronic device, router, switch or other suitable device, or any combination thereof. In at least some feature aspects, at least some of the features or functionalities of the various feature aspects disclosed herein may be implemented in one or more virtual computing environments (eg, network computing clouds, computer systems residing on one or more physical computing machines) virtual machine, or other suitable virtual environment).

Referring now to FIG. 10, a block diagram depicting an exemplary computing device 10 suitable for implementing at least a portion of the features or functions disclosed herein is shown. For example, computing device 10 may be any of the computing machines listed in the preceding paragraphs, or indeed any other electronic device capable of executing software or hardware-based instructions in accordance with one or more programs stored in memory. Computing device 10 may be configured to communicate with various other computing devices such as clients or servers over a communication network such as a wide area network, metropolitan area network, local area network, wireless network, the Internet, or any other network, using known protocols for such communication. Device communication, whether wireless or wired.

In one aspect, computing device 10 includes one or more central processing units (CPUs) 12, one or more interfaces 15, and one or more buses 14 (eg, a peripheral component interconnect (PCI) bus). When acting under the control of appropriate software or firmware, CPU 12 may be responsible for implementing particular functions associated with the functions of a particular configured computing device or machine. For example, in at least one feature aspect, computing device 10 may be configured or designed as a server system utilizing CPU 12 , local memory 11 and/or remote memory 16 and interface 15 . In at least one feature aspect, CPU 12 may cause one or more different types of functions and/or operations to be performed under the control of software modules or components, which may include, for example, an operating system and any suitable application software, drivers, and the like.

CPU 12 may include one or more processors 13, eg, processors from one of the Intel, ARM, Qualcomm, and AMD families of microprocessors. In some feature aspects, processor 13 may include specially designed hardware such as application specific integrated circuits (ASICs), electrically erasable programmable read only memories (EEPROMs), field programmable gates for controlling the operation of computing device 10 Arrays (FPGAs), etc. In certain feature aspects, local memory 11 (such as non-volatile random access memory (RAM) and/or read only memory (ROM), including, for example, one or more levels of cache memory) may also form part of CPU 12 . However, there are many different ways in which memory may be coupled to system 10 . The memory 11 may be used for various purposes, such as, for example, caching and/or storing data, programming instructions, and the like. It should further be appreciated that CPU 12 may be various system-on-chip (SOC) type hardware, which may include additional hardware such as memory or graphics processing chips such as CPUs that are increasingly common in the art, such as QUALCOMM SNAPDRAGON ^™ or Samsung EXYNOS ^™ CPU, such as for mobile devices or integrated devices.

As used herein, the term "processor" is not limited to an integrated circuit referred to in the art involving a processor, mobile processor, or microprocessor, but broadly refers to a microcontroller, microcomputer, programmable logic Controllers, ASICs and any other programmable circuits.

In one aspect, the interfaces 15 are provided as network interface cards (NICs). Generally, NICs control the transmission and reception of data packets over a computer network; for example, other types of interfaces 15 may support other peripherals used with computing device 10. Available interfaces include Ethernet interfaces, frame relay interfaces, cable interfaces, DSL interfaces, token ring interfaces, graphics interfaces, and the like. Additionally, various types of interfaces may be provided, eg, Universal Serial Bus (USB), Serial, Ethernet, FIREWIRE ^™ , THUNDERBOLT ^™ , PCI, Parallel, Radio Frequency (RF), BLUETOOTH ^™ , Near Field Communication (eg, Using Near Field Magnetics), 802.11 (WiFi), Frame Relay, TCP/IP, ISDN, Fast Ethernet Interface, Gigabit Ethernet Interface, Serial ATA (SATA) or External SATA (ESATA) Interface, High Definition Multimedia Interface (HDMI), digital video interface (DVI), analog or digital audio interface, asynchronous transfer mode (ATM) interface, high-speed serial interface (HSSI) interface, point-of-sale (POS) interface, fiber optic data distribution interfaces (FDDIs), etc. . Typically, the interface 15 may include a physical port suitable for communicating with an appropriate medium. In some cases, they may also include separate processors (such as dedicated audio or video processors, which are common in the art for high-fidelity A/V hardware interfaces) and in some cases, volatile and/or non-volatile memory (eg RAM).

Although the system shown in FIG. 10 illustrates one particular architecture of a computing device 10 for implementing one or more aspects described herein, it is by no means at least a portion upon which the features and techniques described herein may be implemented unique device architecture. For example, architectures with one or any number of processors 13 may be used, and such processors 13 may be present in a single device or distributed among any number of devices. In one feature aspect, a single processor 13 handles communications and routing calculations, while in other feature aspects a separate dedicated communications processor may be provided. In various feature aspects, different types of features may be implemented in the system depending on the feature aspects including a client device (eg, a tablet device or smartphone running client software) and a server system (eg, the server system described in more detail below). feature or function.

Regardless of the network device configuration, the system of the features may employ one or more memories or memory modules (eg, remote memory block 16 and local memory 11 ) configured to store data, program instructions, or Additional information (or any combination of the above) relevant to the functionality of the aspects described herein. For example, the program instructions may control the execution of or include the operating system and/or one or more application programs. The memory 16 or the memories 11, 16 may also be configured to store data structures, configuration data, encrypted data, historical system operating information, or any other specific or general non-program information described herein.

Because such information and program instructions can be used to implement one or more of the systems or methods described herein, at least some aspects of the network device can include non-translated machine-readable storage media, which, for example, can be configured or designed to store Program instructions, status information, and the like for performing the various operations described herein. Examples of such non-translational machine-readable storage media include, but are not limited to, magnetic media such as magnetic disks, floppy disks, and magnetic tapes; optical media such as CD-ROM disks; magneto-optical media such as optical disks; Hardware devices, such as read-only memory (ROM), flash memory (common in mobile devices and integrated systems), solid-state drives (SSD), and "hybrid SSD" storage drives, which combine the physical components of solid-state drives and hard-disk drives In a hardware device (which is becoming more common in the art in the case of personal computers), memory, random access memory (RAM), etc. It should be appreciated that such storage devices may be integral and non-removable (eg, RAM hardware modules that may be soldered to the motherboard or otherwise integrated into the electronic device), or they may be removable, such as swappable flash memory modules (such as "thumb drives" or other removable media designed to quickly swap physical storage devices), "hot-swap" hard disk drives or solid state drives, removable optical discs or other such removable media, and such integral storage media and Removable storage media can be used interchangeably. Examples of program instructions include two object code, such as object code generated by a compiler, machine code, such as object code generated by an assembler or linker, byte code, such as object code generated by a JAVA ^TM compiler, which can be virtualized using a Java TM compiler. The interpreter can also be used to execute files containing higher-level code (for example, scripts written in Python, Perl, Ruby, Groovy, or any other scripting language).

In certain feature aspects, the system may be implemented on a separate computing system. Referring now to FIG. 11, a block diagram depicting a typical exemplary architecture of one or more aspects or components thereof on a stand-alone computing system is shown. Computing device 20 includes a processor 21 that can execute software, such as a client application 24, that executes applications that perform one or more functions or aspects. Processor 21 may execute computational instructions under the control of operating system 22, such as, for example, versions of MICROSOFT WINDOWS ^™ operating systems, APPLE macOS ^™ or iOS ^™ operating systems, some Linux operating systems, ANDROID ^™ operating systems, and the like. In many cases, one or more shared services 23 may run in system 20 and may be used to provide common services to client applications 24 . For example, service 23 may be a WINDOWS ^™ service, a user space common service in a Linux environment, or any other type of common service architecture used with operating system 21 . Input device 28 may be of any type suitable for receiving user input, including, for example, a keyboard, touch screen, microphone (eg, for voice input), mouse, touch pad, trackball, or any combination thereof. Output device 27 may be of any type suitable for providing output to one or more users, whether remote or local users of system 20, and may include, for example, one or more screens for visual output, speakers, printers, or any of these. combination. The memory 25 may be random access memory of any structure and architecture known in the art for the processor 21, eg for running software. Storage device 26 may be any magnetic, optical, mechanical, memory, or electrical storage device (eg, those described above, with reference to FIG. 10 ) for storing data in digital form. Examples of storage device 26 include flash memory, magnetic hard disks, CD-ROMs, and the like.

In some feature aspects, the system may be implemented on a distributed computing network, such as a network with any number of clients and/or servers. Referring now to FIG. 12, a block diagram depicting an exemplary architecture 30 for implementing at least a portion of a system according to one aspect over a distributed computing network is shown. According to this characteristic aspect, any number of clients 33 may be provided. Each client 33 may run software for implementing the client portion of the system; the client may include system 20 as shown in FIG. 11 . Furthermore, any number of servers 32 may be provided to handle requests received from one or more clients 33 . The client 33 and the server 32 may communicate with each other via one or more electronic networks 31, which may in various characteristic aspects be the Internet, a wide area network, a mobile telephone network (eg CDMA or GSM cellular network), a wireless network (eg WiFi , WiMAX, LTE, etc.) or a local area network (or indeed any network topology known in the art; no one network topology is preferred in terms of features). Network 31 may be implemented using any known network protocol, including, for example, wired and/or wireless protocols.

Furthermore, in some feature aspects, the server 32 may invoke external services 37 to obtain additional information, or reference additional data regarding a particular call, as needed. For example, external services 37 may be communicated via one or more networks 31 . In various feature aspects, external services 37 may include network-enabled services or functions associated with or installed on the hardware device itself. For example, in one aspect where the client application 24 is implemented on a smartphone or other electronic device, the client application 24 may retrieve a server system 32 stored in the cloud or deployed on one or more premises of a particular enterprise or user. Information on External Services 37.

In some feature aspects, client 33 or server 32 (or both) may utilize one or more dedicated services or devices, which may be deployed locally or remotely over one or more networks 31 . For example, one or more databases 34 may be used or referenced by one or more aspects. It will be understood by those of ordinary skill in the art that the database 34 may be arranged in a variety of architectures and using a variety of data access and manipulation means. For example, in various aspects, one or more of the databases 34 may include relational database systems using Structured Query Language (SQL), while other databases 34 may include alternative data storage technologies such as those known in the art as "NoSQL" (eg , HADOOPCASSANDRA ^™ , GOOGLE BIGTABLE ^™ , etc.). In some feature aspects, various database architectures may be used in accordance with this aspect, such as column-oriented databases, in-memory databases, clustered databases, distributed databases, and even flat-file data stores. It will be understood by those of ordinary skill in the art that any combination of known or future database technologies may be used as appropriate, unless a particular database technology or a particular arrangement of components is specified for a particular aspect described herein. Furthermore, it should be understood that the term "database" as used herein may refer to a physical database machine, a cluster of machines acting as a single database system, or a logical database within an entire database management system. Unless a specific meaning is assigned to a given use of the word "database", it should be interpreted as any of these meanings of the word, all of which are understood by those skilled in the art to be the simple meaning of the word "database".

Similarly, some feature aspects may utilize one or more of the security system 36 and the configuration system 35 . Security and configuration management are common information technology (IT) and network functions, some of which are commonly associated with any IT or network system. It will be understood by those of ordinary skill in the art that any configuration or security subsystem known in the art now or in the future may be used with an aspect without limitation unless the description of any particular aspect specifically requires a particular security 36 or configuration system 35 or method.

FIG. 13 shows an exemplary overview of computer system 40 that may be used in any of various locations throughout the system. It is an example of any computer that can execute code to process data. Various modifications and changes may be made to computer system 40 without departing from the broader scope of the systems and methods disclosed herein. Central processing unit (CPU) 41 is connected to bus 42 , which is also connected to memory 43 , non-volatile memory 44 , display 47 , input/output (I/O) unit 48 and network interface card (NIC) 53 . Typically, I/O unit 48 may be connected to keyboard 49 , pointing device 50 , hard drive 52 and real time clock 51 . The NIC 53 is connected to a network 54, which may be the Internet or a local network, which may or may not have a connection to the Internet. In this example, a power supply unit 45 connected to a mains alternating current (AC) power source 46 is also shown as part of the system 40 . Not shown are possible batteries, as well as many other devices and modifications that are known but not applicable to the particular novel functions of the current systems and methods disclosed herein. It should be understood that some or all of the components shown may be combined together, such as in various integrated applications, such as in Qualcomm or Samsung system-on-chip (SOC) devices, or in combining multiple functions or functions into a single hardware device ( For example, in mobile devices such as smartphones, video game consoles, in-vehicle computer systems such as navigation or car multimedia systems, or other integrated hardware devices).

In various feature aspects, the functionality of a system or method for implementing the various aspects may be distributed among any number of client and/or server components. For example, various software modules may be implemented to perform various functions associated with the system in any particular feature, and these modules may be implemented differently to run on server and/or client components.

Those skilled in the art will appreciate the scope of possible modifications with respect to the various features described above. Accordingly, the invention is defined by the claims and their equivalents.

Claims (10)

1. A system for transfer learning and domain adaptation using a distributable data model, comprising: A distributable model source comprising a memory, a processor and a plurality of programming instructions stored in said memory and executable on said processor, wherein said programmable instructions when executed on said processor make the processor: store multiple machine learning models; generating a distributable model instance based at least in part on the machine learning model; sending the distributable model instance over a network; A transport engine comprising a memory, a processor, and a plurality of programmed instructions stored in said memory and executable on said processor, wherein said programmable instructions, when executed on said processor, cause said processing device: receiving at least a distributable model instance from the model source; and applying a plurality of machine learning algorithms to at least a portion of the received distributable model instances; A directed computational graph engine comprising a memory, a processor, and a plurality of programmed instructions stored in the memory and executable on the processor, wherein the programmable instructions, when executed on the processor, cause the processor: receiving at least a distributable model instance from the model source; creating a second dataset from data stored in the memory based at least in part on transfer learning performed by the transfer engine; training the distributable model instance using the second dataset; and An update report is generated based at least in part on the update to the distributable model instance. 2. The system of claim 1, wherein the machine learning algorithm comprises a probabilistic learning network. 3. The system of claim 2, wherein the probabilistic learning network comprises a Markov logical network. 4. The system of claim 2, wherein the probabilistic learning network comprises a Bayesian network. 5. The system of claim 1, wherein the transport engine: Receive pre-trained data models; Including at least a portion of the pretrained data model into a partially unsupervised machine learning process; and The partially unsupervised machine learning process is applied to the distributable model instance. 6. A method for transfer learning and domain adaptation using a distributable data model, comprising the steps of: (a) storing multiple machine learning models in a distributable model source; (b) generating distributable model instances based at least in part on the machine learning model; (c) sending the distributable model instance via the network; (d) receiving at least a distributable model instance from the distributable model source at the transport engine; (e) applying a plurality of machine learning algorithms to the distributable model instance; (f) using a directed computational graph engine to train the distributable model instance using a plurality of machine learning algorithms executed by the transfer engine; and (g) generating an update report based at least in part on updates to the distributable model instance. 7. The method of claim 6, wherein the machine learning algorithm comprises a probabilistic learning network. 8. The method of claim 7, wherein the probabilistic learning network comprises a Markov logical network. 9. The method of claim 7, wherein the probabilistic learning network comprises a Bayesian network. 10. The method of claim 6, further comprising the step of: Receive pre-trained data models; Including at least a portion of the pretrained data model into a partially unsupervised learning process; and The partially unsupervised machine learning process is applied to the distributable model instance.

Images