Artificial intelligence engine having an architect module

• US 10,733,531 B2
• Filed: 01/26/2017
• Issued: 08/04/2020
• Est. Priority Date: 01/27/2016
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Bonsai AI, Inc. (Microsoft Corporation)
• Inventors: Hammond, Mark Isaac, Browne, Keen McEwan, Campos, Marcos, Brown, Matthew James, Kong, Ruofan, Adams, Megan
• Primary Examiner(s): Hicks, Austin
• Application Number:
  US15/416,904
• Publication Number:
  US 20170213154A1
• Time in Patent Office: 1,286 Days
• CPC Class Codes: G06F 15/80 comprising an array of processing units with common control, e.g. single instruction multiple data processors G06F15/82 takes precedence ; for correlation function computation G06F17/15
• CPC Class Codes: G06F 16/2228 Indexing structures
• CPC Class Codes: G06F 16/951 Indexing; Web crawling techniques
• CPC Class Codes: G06F 18/2148 characterised by the process organisation or structure, e.g. boosting cascade
• CPC Class Codes: G06F 3/03543 Mice or pucks G06F3/03541 takes precedence
• CPC Class Codes: G06F 3/0482 Interaction with lists of selectable items, e.g. menus
• CPC Class Codes: G06F 30/20 Design optimisation, verification or simulation optimisation, verification or simulation of circuit designs G06F30/30
• CPC Class Codes: G06F 30/333 Design for testability [DFT], e.g. scan chain or built-in self-test [BIST]
• CPC Class Codes: G06F 40/166 Editing, e.g. inserting or deleting
• CPC Class Codes: G06F 8/31 Programming languages or programming paradigms
• CPC Class Codes: G06F 8/311 Functional or applicative languages; Rewrite languages
• CPC Class Codes: G06F 8/38 for implementing user interfaces
• CPC Class Codes: G06F 8/437 Type checking
• CPC Class Codes: G06F 9/451 Execution arrangements for user interfaces
• CPC Class Codes: G06F 9/4881 Scheduling strategies for dispatcher, e.g. round robin, multi-level priority queues
• CPC Class Codes: G06F 9/5066 Algorithms for mapping a plurality of inter-dependent sub-tasks onto a plurality of physical CPUs mappping at compile time, see G06F8/451
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/008 based on physical entities controlled by simulated intelligence so as to replicate intelligent life forms, e.g. based on robots replicating pets or humans in their appearance or behaviour
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/08 : Learning methods
• CPC Class Codes: G06N 3/105 : Shells for specifying net layout
• CPC Class Codes: G06N 5/04 : Inference or reasoning models
• CPC Class Codes: G06Q 10/00 : Administration; Management
• CPC Class Codes: H04L 67/01 : Protocols
• CPC Class Codes: H04L 67/02 : based on web technology, e.g. hypertext transfer protocol [HTTP]

Abstract

Provided herein in some embodiments is an artificial intelligence (“AI”) engine hosted on one or more servers configured to cooperate with one or more databases including one or more AI-engine modules. The one or more AI-engine modules include an architect module configured to propose an AI model from an assembly code. The assembly code can be generated from a source code written in a pedagogical programming language describing a mental model of one or more concept modules to be learned by the AI model and curricula of one or more lessons for training the AI model on the one or more concept modules in one or more training cycles. The AI engine can be configured to instantiate a trained AI model based on the one or more concept modules learned by the AI model in the one or more training cycles.

Claims

1. A server system comprising one or more servers configured to host an artificial intelligence (“

AI”

) engine, the AI engine comprising;

one or more AI-engine modules including an architect module,

wherein the architect module is configured to propose an AI model from an assembly code,

wherein the AI engine is configured to train the AI model in one or more training cycles with training data from one or more training data sources,

wherein the assembly code is generated from a source code written in a pedagogical programming language,

wherein the source code includes

a mental model of one or more concept modules to be learned by the AI model using the training data and

curricula of one or more lessons for training the AI model on the one or more concept modules,

wherein the AI engine is configured to instantiate a trained AI model based on the one or more concept modules learned by the AI model in the one or more training cycles,

wherein the AI engine includes a database that stores signatures determined for a plurality of mental models, and stores a plurality of topologies for AI models that have been previously used to solve the plurality of mental models,

wherein the AI engine is configured to determine a signature of the mental model of the source code,

wherein the signatures for the plurality of mental models and the signature for the mental model of the source code are determined using hashing, and

wherein the architect module is configured to select a topology from the plurality of topologies stored in the database based on the determined signature of the mental model of the source code, and instantiate the selected topology for the AI model.

2. The server system of claim 1, the AI engine further comprising:

a compiler configured to generate the assembly code from the source code,

wherein the training data is batched training data, streamed training data, or a combination thereof, and

wherein the AI engine is configured to push or pull the training data from one or more training data sources selected from a simulator, a training data generator, a training data database, or a combination thereof.

3. The server system of claim 2, the AI engine further comprising:

a data stream manager configured to manage streaming of the streamed training data.

4. The server system of claim 1,

wherein the AI engine is configured to operate in a training mode or a predicting mode during the one or more training cycles,

wherein the AI engine in the training mode is configured to

i) instantiate the AI model conforming to the AI model proposed by the architect module and

ii) train the AI model, and

wherein the AI engine in the predicting mode is configured to

i) instantiate and execute the trained AI model on the training data through one or more application programming interface (“

API”

) endpoints for one or more predictions in the predicting mode.

5. The server system of claim 1,

wherein the AI engine is configured to heuristically pick an appropriate learning algorithm from a plurality of machine learning algorithms stored in one or more databases of the server system for training the AI model proposed by the architect module.

6. The server system of claim 5,

wherein the architect module is configured to propose one or more additional AI models, and

wherein the AI engine is configured to

i) heuristically pick an appropriate learning algorithm from the plurality of machine learning algorithms in the one or more databases for each of the one or more additional AI models,

ii) train the AI models in parallel,

wherein the one or more additional AI models are also trained in one or more training cycles with the training data from one or more training data sources, and

iii) instantiate one or more additional trained AI models based on the concept modules learned by the one or more AI models in the one or more training cycles, and

iv) identify a best trained AI model among the trained AI models.

7. The server system of claim 6, the AI engine further comprising:

a trained AI-engine AI model,

wherein the trained AI-engine AI model provides enabling AI for proposing the AI models from the assembly code and picking the appropriate learning algorithms from the plurality of machine learning algorithms in the one or more databases for training the AI models, and

wherein the AI engine is configured to continuously train the trained AI-engine AI model in providing the enabling AI for proposing the AI models and picking the appropriate learning algorithms.

8. The server system of claim 6, the AI engine further comprising:

a meta-learning module configured to keep a record in the one or more databases for

i) the source code processed by the AI engine,

ii) mental models of the source code,

iii) the training data used for training the AI models,

iv) the trained AI models,

v) how quickly the trained AI models were trained to a sufficient level of accuracy, and

vi) how accurate the trained AI models became in making predictions on the training data.

10. The server system of claim 1,

wherein the AI engine is configured to provide one or more training status updates on training the AI model selected from

i) an estimation of a proportion of a training plan completed for the AI model,

ii) an estimation of a completion time for completing the training plan,

iii) the one or more concept modules upon which the AI model is actively training,

iv) mastery of the AI model on learning the one or more concept modules,

v) fine-grained accuracy and performance of the AI model on learning the one or more concept modules, and

vi) overall accuracy and performance of the AI model on learning one or more mental models.

19. The server system of claim 1, wherein the signatures for the plurality of mental models and the mental model of the source code are determined using hashing such that mental models that have similar machine learning algorithmic characteristics have similar signatures.

9. A server system comprising one or more servers configured to host an artificial intelligence (“

AI”

) engine, the AI engine comprising;

one or more AI-engine modules including an architect module,

wherein the architect module is configured to propose an AI model from an assembly code,

wherein the AI engine is configured to train the AI model in one or more training cycles with training data from one or more training data sources,

wherein the assembly code is generated from a source code written in a pedagogical programming language,

wherein the source code includes;

a mental model of one or more concept modules to be learned by the AI model using the training data and

curricula of one or more lessons for training the AI model on the one or more concept modules, each of the one or more lessons configured to optionally use a different flow of the training data,

wherein the AI engine is configured to instantiate a trained AI model based on the one or more concept modules learned by the AI model in the one or more training cycles,

wherein the architect module is configured to instantiate a topology of the AI model,

wherein the AI engine is configured to make determinations regarding

i) when to train the AI model on each of the one or more concept modules and

ii) how extensively to train the AI model on each of the one or more concept modules, and

wherein the determinations are based on the relevance of each of the one or more concept modules in one or more predictions of the trained AI model based upon the training data.

11. A method comprising:

at a processor of a computer device configured to host an artificial intelligence (“

AI”

) engine;

proposing an AI model,

wherein the AI engine includes one or more AI-engine modules including an architect module for proposing the AI model from an assembly code;

training the AI model,

wherein the AI engine is configured to train the AI model in one or more training cycles with training data from one or more training data sources;

compiling the assembly code from a source code,

wherein a compiler is configured to generate the assembly code from the source code written in a pedagogical programming language,

wherein the source code includes

a mental model of one or more concept modules to be learned by the AI model using the training data and

curricula of one or more lessons for training the AI model on the one or more concept modules, each of the one or more lessons configured to optionally use a different flow of the training data;

instantiating a trained AI model,

wherein the AI engine is configured for instantiating the trained AI model based on the one or more concept modules learned by the AI model in the one or more training cycles; and

heuristically picking an appropriate learning algorithm,

wherein the AI engine is configured for picking the appropriate learning algorithm from a plurality of machine learning algorithms stored in one or more databases for training the AI model proposed by the architect module.

12. The method of claim 11, further comprising:

pushing or pulling the training data,

wherein the AI engine is configured for pushing or pulling the training data from one or more training sources selected from a simulator, a training data generator, a training data database, or a combination thereof, and

wherein the training data is batched training data, streamed training data, or a combination thereof.

13. The method of claim 11, further comprising:

operating the AI engine in a training mode or a predicting mode during the one or more training cycles,

wherein the AI engine in the training mode

i) instantiates the AI model conforming to the AI model proposed by the architect module and

ii) trains the AI model, and

wherein the AI engine in the predicting mode

i) instantiates and executes the trained AI model on the training data through one or more application programming interface (“

API”

) endpoints for one or more predictions in the predicting mode.

14. The method of claim 11, further comprising:

proposing one or more additional AI models,

wherein the architect module is configured for proposing the one or more additional AI models;

heuristically picking an appropriate learning algorithm from the plurality of machine learning algorithms stored in the one or more databases with the AI engine for each of the one or more additional AI models;

training the AI models in parallel with the AI engine,

wherein the one or more additional AI models are also trained in one or more training cycles with the training data from one or more training data sources;

instantiating one or more additional trained AI models with the AI engine based on the concept modules learned by the one or more AI models in the one or more training cycles; and

identifying a best trained AI model among the trained AI models with the AI engine.

15. The method of claim 14, further comprising:

providing enabling AI for proposing the AI models from the assembly code and picking the appropriate learning algorithms from the plurality of machine learning algorithms in the one or more databases for training the AI models; and

continuously training a trained AI-engine AI model with the AI engine to provide the enabling AI for proposing the AI models and picking the appropriate learning algorithms.

16. The method of claim 14, further comprising:

keeping a record in the one or more databases with a meta-learning module,

wherein the record includes

i) the source code processed by the AI engine,

ii) mental models of the source code,

iii) the training data used for training the AI models,

iv) the trained AI models,

v) how quickly the trained AI models were trained to a sufficient level of accuracy, and

vi) how accurate the trained AI models became in making predictions on the training data.

17. The method of claim 11, further comprising:

making determinations with the AI engine regarding

i) when to train the AI model on each of the one or more concept modules and

ii) how extensively to train the AI model on each of the one or more concept modules,

wherein the determinations are based on the relevance of each of the one or more concept modules in one or more predictions of the trained AI model based upon the training data.

18. The method of claim 11, further comprising:

providing one or more training status updates with the AI engine on training the AI model,

wherein the one or more training status updates are selected from

i) an estimation of a proportion of a training plan completed for the AI model,

ii) an estimation of a completion time for completing the training plan,

iii) the one or more concept modules upon which the AI model is actively training,

iv) mastery of the AI model on learning the one or more concept modules,

v) fine-grained accuracy and performance of the AI model on learning the one or more concept modules, and

vi) overall accuracy and performance of the AI model on learning one or more mental models.