Granted Patent B2

US 12,268,187 · App. 17/647,063 · Granted Apr 8, 2025

Inventors: Jenna Lau Caruso (Holland Landing, CA); Samaya Madhavan (Euless, TX); Stephen Haertel (Ajax, CA); Geetha Kola (Bangalore, IN); Laura Jane Bennett (Alameda, CA); Lior Aronovich (Thornhill, CA)

Assignee: International Business Machines Corporation

A01K15/021G06N20/00

Abstract

Claims (98)

1. A method for automatic animal behavior training, comprising, using a computer processor:

receiving user configuration information and user feedback information from a user regarding a specific animal;

receiving input from sensors for the specific animal;

determining a mapping between the input and a set of positive and negative training actions to apply for the specific animal;

in a generating phase:

generating a training set for an animal behavior machine learning model (MLM) based on sensed animal behavior patterns and the user feedback information;

in an MLM training phase:

training the animal behavior MLM based on the training set using back propagation that includes gradient descent to reduce an objective function by altering a set of weights and biases of the animal behavior MLM to increase accuracy of the animal behavior MLM, wherein the animal behavior MLM comprises a neural network;

in an application phase:

sensing, by the sensors associated with the specific animal, a behavior of the specific animal to produce a sensed behavior; and

conditioned upon the sensed behavior, providing an animal training action,

wherein the user configuration information and the user feedback information comprise:

specific animal name, type, breed, age, and weight;

which of the sensors to use for the specific animal;

which training devices to use for the specific animal;

a user mapping of training action types and training action levels for training action devices to corresponding behaviors of the specific animal; and

static definitions of specific sensor data mappings to specific training action types and specific training action levels.

2. The method of claim 1 further comprising initializing the animal behavior MLM by starting from a pre-calculated model instance that is associated with an attribute selected from the group consisting of the specific animal's: type, breed, age, and weight.

3. The method of claim 2 , further comprising initializing a plurality of animal behavior MLMs that include the animal behavior MLM, by starting from a plurality of pre-calculated model instances that are associated with a plurality of the specific animal's: type, breed, age, and weight.

4. The method of claim 1 , wherein the generating of the training set further comprises receiving the user feedback information regarding an animal action type and an animal action level to take for a sensed animal behavior pattern of the specific animal.

5. The method of claim 4 , wherein the generating of the training set further comprises, to prepare an input record for the training set from sensor input data of the sensors, an operation of:

automatically determining a start time from which to include the sensor input data in a training record conditioned upon the user not indicating the start time, by using at least one of the following operations:

in each sensor input, searching back for a value which has a larger relative difference from its preceding value above a threshold;

in each sensor input, searching back for an outlier value;

taking a value with a largest difference from its preceding value conditioned upon no outlier being found;

using the found value to delimit and specify the start time with regards to this sensor data;

limiting the searching back to a time window; and

selecting a time of an oldest extreme value in time windows across all the sensors as the start time for the training record.

6. The method of claim 5 , wherein a different time length is used for each of the sensors.

7. The method of claim 6 , wherein the generating of the training set further comprises, to prepare an input record for the training set from the input sensors data:

using, for each time length of a respective sensor of the sensors, a separate set of animal behavior MLMs.

8. The method of claim 4 , wherein the generating of the training set further comprises, to prepare an input record for the training set from the sensor data, an operation selected from the group consisting of:

aggregating multiple readings of each sensor into a representative value;

using at least one of an aggregation and selection operation that is appropriate for each sensor;

including multiple values from a sensor in a generated training record;

applying feature extraction techniques for sensor data that includes multiple features;

removing records that contain outliers in one or more of their sensor record fields;

applying dimensionality reduction on the records of a training set to select a most efficient sub-set of dimensions to use for training; and

applying at least one of training set partitioning, sharding, and slicing technique to increase at least one of accuracy and resiliency of the animal behavior MLM.

9. The method of claim 4 , wherein the user feedback information further comprises a classification of the sensed animal behavior pattern.

10. The method of claim 1 , wherein validating of the animal behavior MLM further comprises:

providing predictions to the user based on the specific animal's behavior patterns.

11. The method of claim 1 , wherein:

providing system generated classifications utilizes results of the animal behavior MLM;

the method further comprising:

using the animal behavior MLM to check readings of the sensors;

aggregating results of various types or instances of animal behavior MLMs;

performing a user action selected from the group consisting of maintaining, removing, and modifying any sub-set of animal behavior MLMs based on user input; and

notifying the user of entering the application phase.

12. The method of claim 1 , wherein the animal training action is taken:

based on system predictions in the application phase.

13. The method of claim 1 , wherein the user feedback information comprises a classification of an animal behavior pattern and associated requested animal training action.

14. The method of claim 1 , wherein the animal behavior MLM comprises the neural network that contains algorithms and data to perform machine learning functionality.

15. The method of claim 1 , where pre-existing training data is also used as a basis when generating the training set for the animal behavior MLM.

16. The method of claim 1 , further comprising:

capturing a behavior pattern of the specific animal, wherein the user feedback information identifies how to classify the behavior pattern.

17. The method of claim 1 , further comprising:

in a validation phase,

validating the animal behavior MLM by providing system generated classifications.

18. The method of claim 1 , wherein the animal training action includes a training action type and a training action level of a training action device for the specific animal.

19. A computer program product for automatic animal behavior training, comprising:

a computer-readable storage medium having program instructions embodied thereon to perform steps of:

receiving user configuration information and user feedback information from a user regarding a specific animal;

receiving input from sensors for the specific animal;

determining a mapping between the input and a set of positive and negative training actions to apply for the specific animal;

in a generating phase:

generating a training set for an animal behavior machine learning model (MLM) based on sensed animal behavior patterns and the user feedback information;

in an MLM training phase:

training the animal behavior MLM based on the training set using back propagation that includes gradient descent to reduce an objective function by altering a set of weights and biases of the animal behavior MLM to increase accuracy of the animal behavior MLM, wherein the animal behavior MLM comprises a neural network;

in an application phase:

sensing, by the sensors associated with the specific animal, a behavior of the specific animal to produce a sensed behavior; and

conditioned upon the sensed behavior, providing an animal training action,

wherein the user configuration information and the user feedback information comprise:

specific animal name, type, breed, age, and weight;

which of the sensors to use for the specific animal;

which training devices to use for the specific animal;

a user mapping of training action types and training action levels for training action devices to corresponding behaviors of the specific animal; and

static definitions of specific sensor data mappings to specific training action types and specific training action levels.

20. A computer system for automatic animal behavior training, comprising:

a processor;

a computer-readable storage medium; and

program instructions stored on the computer-readable storage medium to cause the processor to perform operations comprising:

receiving user configuration information and user feedback information from a user regarding a specific animal;

receiving input from sensors for the specific animal;

determining a mapping between the input and a set of positive and negative training actions to apply for the specific animal;

in a generating phase:

generating a training set for an animal behavior machine learning model (MLM) based on sensed animal behavior patterns and the user feedback information;

in an MLM training phase:

training the animal behavior MLM based on the training set using back propagation that includes gradient descent to reduce an objective function by altering a set of weights and biases of the animal behavior MLM to increase accuracy of the animal behavior MLM, wherein the animal behavior MLM comprises a neural network;

in an application phase:

sensing, by the sensors associated with the specific animal, a behavior of the specific animal to produce a sensed behavior; and

conditioned upon the sensed behavior, providing an animal training action,

wherein the user configuration information and the user feedback information comprise:

specific animal name, type, breed, age, and weight;

which of the sensors to use for the specific animal;

which training devices to use for the specific animal;

a user mapping of training action types and training action levels for training action devices to corresponding behaviors of the specific animal; and

static definitions of specific sensor data mappings to specific training action types and specific training action levels.

Assignments (1)

Continuity (1)

References Cited (38)

Cited By (2)

US 12,490,717 US 12,718,524