Method and apparatus for managing a trained computer model

Abstract

A method for providing a computer model as a tradable asset is described. Prepare a trained computer model. This includes developing and training a computer model using training data and determining acceptance criteria for the computer model such that when the acceptance criteria are met, the computer model is determined to be a trained computer model. The trained computer model is packaged for use by a third party and stored securely. Establish a token corresponding to the trained computer model. Post the token and token transactions to a blockchain to adapt the token for use as a tradable asset. Access to the trained computer model is provided to a third party that has obtained the right to use the trained computer model by obtaining the rights to the token. A related system for developing a computer model and providing the computer model as a tradable asset is also described.

Description

Technical Field

[0001] The present invention relates to a method and apparatus for managing trained computer models.

Background Art

[0002] Machine learning (ML) has increasingly been used to address a very wide range of problems in real-world environments. A number of algorithmic approaches are available for the development of computer models based on training data. The development of a model typically involves the use of one or more algorithms on training data in an iterative model development process until a model that meets the desired effectiveness criteria is developed.

[0003] In more or less any real-world situation, this process is not entirely automated. The role of a human application domain expert is actually important at least in determining the effectiveness criteria, but often also in shaping the overall model development process (e.g., by determining how to evaluate which variables). The "domain" here can be considered in a very broad sense and can be applied to any problem context where real-world knowledge is required to appropriately evaluate the problem or its solution.

[0004] The development of computer models using best practices in machine learning generally requires specialized tools and techniques. Specialized tools and techniques are generally not accessible or controllable by domain experts in most application areas. For such domain experts, there is often little incentive to participate in the development of computer models. Experts face the prospect of developing important expertise outside of their own domain to develop the model or providing knowledge to model developers without a clear path to receive compensation if the resulting computer model is proven to have significant value. It is desirable to have a process for computer model development that enables and rewards easy interaction with experts in fields other than machine learning.

Summary of the Invention

Problems to be Solved by the Invention

[0005] The present invention has been conceived against such a background.

Means for Solving the Problems

[0006] In a first aspect, the present invention provides a method of offering a computer model as a tradable asset, comprising: developing and training a computer model using training data; determining acceptance criteria for the computer model such that when the acceptance criteria are met, the computer model is determined to be the trained computer model; preparing the trained computer model; packaging the trained computer model for use by a third party and storing the trained computer model securely; establishing a token corresponding to the trained computer model; posting the token and token transactions to a blockchain so as to adapt the token for use as a tradable asset; and providing access to the trained computer model to a third party that has obtained the right to use the trained computer model by obtaining the rights to the token.

[0007] By using this approach, it is possible to appropriately protect and reward the contributions of model developers. Providing the trained computer model in a securely stored form and granting access to those who have obtained rights through tokens are important elements of the technical infrastructure for achieving this goal.

[0008] In an embodiment, the trained computer model is a machine learning model. The process of preparing the trained computer model may comprise repeating the processes of feature selection, algorithm selection, model construction, and model testing until the acceptance criteria are met.

[0009] This approach to preparing the trained computer model enables effective interaction, for example, between domain experts who may play an important role in determining the acceptance criteria and the model preparation system.

[0010] Packaging a trained computer model may further comprise the model creator digitally signing the trained computer model. Storing the trained computer model securely may comprise storing the trained computer model in an encrypted form encrypted by a key managed by the model creator or model owner. Establishing a token may comprise applying a hash function to the trained computer model and signing the hash result with the model creator's private key.

[0011] In an embodiment, providing access to a trained computer model to a third party may comprise establishing a shared secret between the third party and the model owner or model creator and encrypting the means of access to the trained computer model using the shared secret. Such a shared secret may be established using Diffie-Hellman key exchange.

[0012] In a second aspect, the present invention provides a system for developing a computer model and providing the computer model as a tradable asset, the system comprising a model creation module for developing and training a computer model using training data and adapting the user so that it is determined that the computer model meets a predetermined acceptance criterion to create a trained computer model; a model packaging module for packaging the trained computer model for use by a third party, securely storing the trained computer model, and establishing a token corresponding to the trained computer model; and a model management module for posting the token and the token transaction to a blockchain so as to adapt the token for use as a tradable asset and providing access to the trained computer model to a third party that has obtained the right to use the trained computer model by obtaining the right to the token.

Brief Description of the Drawings

[0013] For a better understanding of the present invention, its preferred non-limiting embodiments will be described by way of example only with reference to the accompanying drawings.

[0014]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

MODE FOR CARRYING OUT THE INVENTION

[0015] Figure 1 shows an infrastructure that may be adopted in an embodiment of the present invention. Here, the model creator 1 who uses the computer device 1a communicates with the model development and management service 4 shown as a service accessed via the cloud 5 or other network connections. This service 4 is of course implemented in a suitable computer architecture, particularly in a suitably programmed processor that interacts with memory. Three modules are shown within the model development and management service 4: a model creation tool 41, a model packaging tool 42, and a model management service 43. When the model is created, it is stored in a secure model storage 6. The rights of the model are managed using a blockchain 7. In this way, the model owner 2 (communicating via the computer device 2a) and the model user 3 (communicating via the computer device 3a) form part of an ecosystem effective for the development and use of the model. Another party to this ecosystem is the model trainer 8 who uses their own computer device 8a. The various elements of this ecosystem will be described in more detail below.

[0016] Figure 1 shows four different human roles. These roles have the following meanings.

[0017] The model creator 1 (MC) is the creator of a set of intermediate products that collectively represent the model. Individual elements (such as a specific algorithm) may be created by others, but the set is the responsibility of the model creator 1. This role does not change over time, although others (such as the model owner 2) may change.

[0018] The model owner 2 (MO) is the entity that holds the legal ownership of the model. As will be described later, since the model is a transferable commodity, different entities may assume this role over time.

[0019] The model user 3 (MU) is simply the party that uses the model and typically obtains the right to do so from the model owner 2 or an agent or service acting on behalf of the model owner. In practice, the model user 3 may be a software component licensed to use the model.

[0020] The model trainer 8 (MT) can further train an already developed model. The services of the model trainer 8 may be required by the model user 3 to adapt the model to a specific user's problem or the user's situation. The model trainer 8 may have expertise in model development or expertise on the subject matter (which may also be obtained from the model creator 1). It is assumed that the model trainer 8 does not have the right to the derivative works created by it.

[0021] Note that one entity may play multiple roles. For example, the model creator 1 may be the first model owner 2 or the model creator 1 may play the role of the model trainer 8.

[0022] In an embodiment of the present invention, the following processes are performed in the ecosystem. The model creator 1 first creates a model using the model creation tool 41. When the model is created, the model is packaged using the model packaging tool 42 and stored in the secure model storage 6 in a securely packaged form. Further, the model packaging tool 42 can create a tokenized form of the packaged model (referred to herein as a non-fungible model by way of example of a non-fungible token) and manage it as a tradable asset through the model management service 43. The model management service 43 is supported by a blockchain 7 that stores information regarding the ownership and rights of the packaged model in an immutable manner. The model management service 43 supports the trading of models in various ways.

[0023] Next, these processes will be described in detail. Figure 2 shows the steps involved in training the model and establishing the packaged model. A model, in the broadest sense, is a set of algorithms implemented using one or more computer languages, and the trained model has parameters developed through training using training data (typically, real-world data or data extrapolated from real-world data). The present invention does not require the use of a specific algorithm and may use a conventional machine learning training process. None of these are fundamental to the present invention; rather, it relates to methods for capturing and protecting the expertise used in the model development and training processes and the resulting methods for storing and using the trained model.

[0024] As a result, the process of model training is basically conventional and typically requires the following steps. · Generation and retention of relevant, clean, and consistent data · Selection of training data from available data · Use of test data to validate the model · Provision of training data appropriately prepared for the machine learning pipeline · Execution of the machine learning pipeline and result verification · Judgment of the effectiveness of the model against test data (which may be labeled in some cases)

[0025] These steps typically require the knowledge of computer scientists, data scientists, and domain experts. Untrained models may primarily be the product of expertise in computer science and data science. Typically, this requires the development of new algorithms and the curation of old algorithms, the development of appropriate execution pipelines, and the establishment of a validation process. In embodiments such as those shown in FIG. 1, this is provided as a general-purpose process and toolkit from which a trained model can be derived. In contrast, trained models use the expertise of all three types of experts, but generally, in domain-specific models, the role of the domain expert is important. In embodiments of the present invention, to enable the development of a trained model, which is the essential outcome of the domain expert who makes the choices and decisions that determine the form of the trained model, it is contemplated that the trained model may be provided through an interface through which the domain expert can interact. The trained model embodying the domain expert's expertise thus becomes an asset that can be packaged, protected, and traded.

[0026] It should be noted that the above specific use cases are those where the selection and determination are made solely by the domain expert, but this is not necessarily the case. The "model creator" may be a team or another form of expert (e.g., a data scientist) who has obtained the required level of domain knowledge from another source.

[0027] Figure 2 shows the model creation process. As would be understood by those skilled in the art, the steps performed here are widely common in machine learning-based model development, and it is not necessary to explain each step in detail here. The first step is to establish (200) a dataset for use in the training process for both training and testing the model, which needs to be appropriately curated such that the dataset is clean and representative of the system being modeled. Thereafter, the features to be used in the model are selected (210), which do not have to be all the features of the dataset, and the algorithm of the model is selected or constructed (220) from the available algorithms. Thereafter, the features used in the model are used to build (230) the entire model using the training data from the dataset, and then the entire model is tested (240) using the test data from the dataset. Thereafter, the test results are evaluated (250) to determine whether they meet the criteria for the effectiveness of the model. This is where the contribution of domain experts is usually particularly important, and if the test results meet these criteria, an effective model has been developed, and the process can proceed to the packaging (260) of the model, the protection (270) of the model, and the establishment (280) of tokens. If these criteria are not met, the process returns to the feature selection step 210 and is repeated until the effectiveness criteria of the model are met (or until the process is abandoned).

[0028] The packaging step 260 and subsequent steps are shown in more detail in FIG. 3. When the validity requirements are met, a trained model is generated. In principle, this can be used in a trained real-world situation. The trained model is, at this point, a collection of intermediate products (metadata, binaries, scripts, static libraries, and other intermediate products specific to the model "container") that can be deployed by the model user. The collection of intermediate products may be adopted to be automatically deployed, for example, when loaded onto the model user device, or may be adapted to be deployed in a model deployment application of the model user computer device. These intermediate products organized within the container are the packaged model. This form of trained model is deployable but not secure.

[0029] To make the model in a secure form, it is necessary to encrypt the model and hold it in a way that can be accessed by the model user. These steps are performed here by the model packaging tool 42. The step of protecting the package (270) can be performed using conventional asymmetric cryptography using a public key infrastructure. To enable the container containing the intermediate products of the trained model to be authenticated by any party with access to the public key of the model creator, the container containing the product of the trained model can be signed with the private key of the model creator. The packaged model may be stored in an essentially secure storage 6, or the packaged model may be stored in an encrypted form within the secure storage 6 to enable only the model creator or their agent to access the packaged model (e.g., by encryption with the public key managed by the model creator with the private key).

[0030] The product of this step is the Safe Trained Model (STM) itself, but as will be explained later, a further safety process is required to make the STM usable by model users. At this point, a Non-Fungible Model (NFM), which is a digital asset related to the Safe Trained Model, is also created (280). This requires a digital identity indicating that it is a feature unique to the NFM, which can be achieved by applying a hash function (any conventional hash function suitable for the amount of data provided may be used.) to the result signed by the private key of the model creator. Such a digital representation of the STM is stored in the blockchain 7 and used for the establishment (and transactions) of ownership and usage rights.

[0031] Before discussing the use of NFM, we will further examine the security model in relation to Figure 4 and consider how a secure trained model is first created and stored, and how a secure trained model is provided to the model user (in this discussion, we assume that the model user already has the right to use the model under the condition that the agreed conditions are met). Model creator 1 establishes (410) a key pair for use in relation to use by model user 3 (this may be specific to a particular creator / user interaction and, in some cases, may be diversified from a master key pair). The same applies to model user 3 (420). At this point, the insecure model intermediate products are aggregated (430) as described above and packaged (440) for storage in a version that is signed by model creator 1 and securely stored with the associated NFM. When it is determined that model user 3 has the right to use the STM, a package for that model user to use needs to be developed. This is typically generated by using the model creator key and the model user's key pair. A standard way to do this is to establish a shared secret using Diffie-Hellman key exchange. In this case, this shared secret can be used to encrypt (450) the STM into a deployable package 460. This can also be done in an embodiment, for example, in a way that there is no risk of security breach by a malicious model user if the model user is permitted access to a service interface where the model can be executed.

[0032] Here, it should be noted that the role of model creator 1 can be assumed by model owner 2. Model creator 1 may not be responsible for more than establishing the use of the model so that it can be appropriately compensated for such use after the initial creation. Next, we will further examine the model management service 43 in relation to Figure 5.

[0033] In FIG. 5, two processes are shown. The first process is the transfer of ownership of the STM to the model owner, which is achieved by transferring the NFM to the model owner. First, model creator 1 adds the NFM to blockchain 7. This can be done in a conventional way using a conventional blockchain architecture. The NFM is stored in the blockchain along with the relevant address indicating the address to go to in order to obtain the rights of the STM (510) (the identification of the model creator may or may not be present, but this is not essential), or the details of the STM are established so that the candidate model owner or user can establish whether they wish to obtain the rights of the model. Then, the candidate model owner can make a "purchase" request (520) regarding the model, which is recorded on blockchain 7. The model creator can either accept or reject this request. Here, it is assumed that the request has been accepted, and a "sale" notification 530 is also stored on blockchain 7, and any financial arrangements between model creator 1 and model owner 2 are established in parallel (all of these operations can be performed through model management service 43). Then, the appropriate rights to manage the STM are transferred to the new model owner 2 along with the ownership of the NFM (540). Since all stages of the ownership transfer are recorded on blockchain 7, the validity of the transaction can be proven.

[0034] Another process shown in FIG. 5 is the release of the NFM to enable the model user 3 to use the STM. Here, this is shown as a process directly between the model user 3 and the model owner 2, which is possible, for example, if the address details of the model owner are stored in the blockchain 7 and will not be recorded on the blockchain, but in other embodiments, all these steps may be recorded on the blockchain 7. The candidate model user 3 makes a request 560 to release the NFM to the model owner 2. If this is accepted, the model owner 2 makes an offer (570), and if the contract is concluded, the model owner 2 cooperates with the model user 3 to execute the steps 580 shown in FIG. 4 to prepare the STM in a form that the model user 3 can use. Thereafter, the model user 3 can use the STM within the lease conditions until the lease expires 590.

[0035] By accessing the STM through the creation of the NFM, an effective market for the trained model can be developed. The NFM typically has the following features or related advantages. · Each of the NFMs is digitally unique and cannot be copied conventionally, functions as a proxy for the STM, and cannot be used unless usage is permitted through the NFM. · The ownership of all NFMs can be tracked and verified. · The NFM (and related digital assets) can be traded in various possible trading models as digital intermediate products (which may be 1-1 or 1-many). · The market for NFMs typically has only technical constraints and can relax the technical constraints when more technical resources become available. · The model creator has the freedom to decide whether to retain or trade the ownership. Generally, all contract options are available.

[0036] The actual use of the model trained by the model user may be in a completely conventional manner, which need not be further explained here, but it should be noted that there may be situations where it is desirable to permit other parties to access the STM. One such party may be the model trainer 8 as shown in FIG. 1 which may be needed when the STM needs to adapt to or be retrained in the model user's situation. To permit access by the model trainer, the same process as shown in FIG. 4 for permitting access by the model user 3 can be used to permit the model trainer 8 to access the STM. If the need for the model trainer is known from the beginning, three-party Diffie-Hellman can be used to give the model trainer access rights. Alternatively, the model owner may simply interact with the model trainer to permit access by the model trainer.

[0037] As indicated here, by the operation of the model management service 43, each of the parties can interact with the model management service 43 to manage rights, obligations, and rewards. The model creator 1 can establish the use of the model and establish the benefits determined between the model creator 1 and the model owner 2 based on the use. The model owner can establish the use (and thus the revenue) and manage how the NFM (and thus the STM) is leased (for what purpose and for what period). The model user 3 can determine what NFM can be used under what conditions.

[0038] As can be understood by those skilled in the art, other embodiments can be provided within the spirit and scope of the present invention described herein.

Claims

1. A method of providing a computer model as a tradable asset, comprising: developing and training a computer model using training data, and determining acceptance criteria for the computer model such that when the acceptance criteria are met, the computer model is determined to be a trained computer model; and packaging the trained computer model for use by a third party and securely storing the trained computer model; establishing a token corresponding to the trained computer model; posting the token and transactions of the token to a blockchain so that the token is adapted for use as a tradable asset; providing access to the trained computer model to a third party that has obtained the right to use the trained computer model by obtaining the right to the token; A method comprising the above steps.

2. The method according to claim 1, wherein the trained computer model is a machine learning model.

3. The method according to claim 1 or claim 2, wherein preparing the trained computer model comprises repeating processes of feature selection, algorithm selection, model construction, and model testing until the acceptance criteria are met.

4. The method according to any one of claims 1 to 3, wherein packaging the trained computer model further comprises digitally signing the trained computer model by a model creator.

5. The method according to any one of claims 1 to 4, wherein securely storing the trained computer model comprises storing the trained computer model in an encrypted form encrypted by a key controlled by a model creator or a model owner.

6. The method according to any one of claims 1 to 5, wherein establishing the token comprises applying a hash function to the trained computer model and signing the hash result with a private key of the model creator.

7. Providing access to the trained computer model to the third party comprises establishing a shared secret between the third party and the model owner or model creator, and encrypting the means of access to the trained computer model using the shared secret. The method according to any one of claims 1 to 6.

8. The method according to claim 7, wherein the shared secret is established using Diffie-Hellman key exchange.

9. A system for developing a computer model and providing the computer model as a tradable asset, a model creation module that develops and trains a computer model using training data and creates a trained computer model adapted to a user to determine that the computer model meets a predetermined acceptance criterion; a model packaging module that packages the trained computer model for use by a third party, securely stores the trained computer model, and establishes a token corresponding to the trained computer model; a model management module that posts the token and the token transaction to a blockchain so as to adapt the token for use as a tradable asset, and provides access to the trained computer model to a third party that has obtained the right to use the trained computer model by obtaining the right of the token. A system comprising: ​

Images