# US 2025/0156762 A1 · Energy efficient machine learning on the edge with query-based knowledge assistance

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2025/0156762 A1
- Document ID: US20250156762A1
- Status: Published application
- Kind: A1
- Application number: 18/944,744
- Filed: 2024-11-12
- Published: 2025-05-15
- Earliest claimed: 2023-11-13
- Assignees: SRI International
- Applicants: SRI International
- Inventors: Aswin NADAMUNI RAGHAVAN; David Chao ZHANG; Saurabh FARKYA; Zachary Alan DANIELS; Michael PIACENTINO; Gooitzen S. VAN DER WAL; Philip MILLER; Michael Richard LOMNITZ; Abrar Abdullah RAHMAN; Edison MUCLLARI; Muhammad Shahir RAHMAN
- Web page: https://patentgrep.com/patent/US20250156762A1

## CPC

- G06N 20/00: Machine learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A method, apparatus, and system for efficient machine learning with query-based knowledge assistance includes determining a state of data captured by a sensor in communication with a first edge device to determine if the captured data includes data that is out of distribution based on a trained inference model of the first edge device, if it is identified that an amount of out of distribution data in the captured data is preventing the trained inference model from making an accurate prediction, communicating a request for resources to a second edge device or a server to elicit a response from the second edge device or the server including resources required to update the trained inference model, receiving the requested resources, updating the trained inference model using the received resources, and making a prediction for the received captured data using the updated, trained inference model.

## Claims

### Claim 1

A method for efficient machine learning with query-based knowledge assistance on a first edge device, comprising:
receiving data captured by at least one sensor in communication with the first edge device;
determining a state of the captured data to determine if the captured data includes data that is out of distribution based on a trained inference model of the first edge device;
if the determined state identifies that an amount of out of distribution data in the captured data is preventing the trained inference model from making an accurate prediction from the captured data, determining a request for resources to be communicated to at least one of a second edge device or a server;
communicating the request for resources to the at least one of the second edge device or the server to elicit a response from the at least one of the second edge device or the server including resources required to update the trained inference model to enable the updated trained inference model to make an accurate prediction from the captured data;
receiving the requested resources;
updating the trained inference model using the received resources to enable the updated trained inference model to make an accurate prediction from the captured data; and
making a prediction for the received captured data using the updated, trained inference model.

### Claim 2 (depends on 1)

The method of claim 1, further comprising identifying, in the request for resources, what resources are required to enable the trained inference model to make an accurate prediction from the captured data;
wherein the required resources are determined using a second machine learning model of the first edge device.

### Claim 3 (depends on 1)

The method of claim 1, wherein the request for resources enables the at least one of the second edge device or the server to determine what resources are required to enable the trained inference model to make an accurate prediction from the captured data, using a second learning model.

### Claim 4 (depends on 3)

The method of claim 3, wherein the request for resources comprises at least information regarding the trained inference model and information regarding the captured data.

### Claim 5 (depends on 1)

The method of claim 1, further comprising:
adjusting capture parameters of at least one of the at least one sensor to capture additional resources for updating the trained inference model.

### Claim 6 (depends on 1)

The method of claim 1, wherein the server retrieves the required resources from a database comprising heterogeneous models including at least one large model and at least one smaller model, and wherein the database communicates resources to the server based on an available bandwidth between the database and the server.

### Claim 7 (depends on 1)

The method of claim 1, wherein the first edge device retrieves the required resources from a network of heterogenous edge devices including at least one large model and at least one smaller model, and wherein at least one edge device of the network of edge devices communicates resources to the first edge device based on resource constraints of at least one of the first edge device or the edge devices of the network.

### Claim 8 (depends on 6)

The method of claim 6, wherein the at least one large model and at least one smaller model are independently trained without cross training.

### Claim 9 (depends on 7)

The method of claim 7, wherein the at least one large model and at least one smaller model are independently trained without cross training.

### Claim 10 (depends on 1)

The method of claim 1, further comprising:
providing reinforcement learning by monitoring a prediction of the trained inference model after an update and providing a reward to at least one of the first edge device, the second edge device or the server based on a result of the prediction of the trained inference model.

### Claim 11 (depends on 1)

The method of claim 1, wherein the request for the required resources is communicated to the server based on a bandwidth available between the first edge device and at least one of the second edge device or the server at the time of the request.

### Claim 12 (depends on 11)

The method of claim 11, further comprising:
providing reinforcement learning by monitoring communications between the first edge device and at least one of the second edge device or the server and providing a respective reward to at least one of the first edge device, the second edge device or the server based on an amount of bandwidth used for at least each communication request for the required resources.

### Claim 13

An apparatus for efficient machine learning with query-based knowledge assistance on a first edge device, comprising:
a processor; and
a memory accessible to the processor, the memory having stored therein at least one of programs or instructions executable by the processor to configure the apparatus to:
receive data captured by at least one sensor in communication with the first edge device;
determine a state of the captured data to determine if the captured data includes data that is out of distribution based on a trained inference model of the first edge device;
if the determined state identifies that an amount of out of distribution data in the captured data is preventing the trained inference model from making an accurate prediction from the captured data, determine a request for resources to be communicated to at least one of a second edge device or a server;
communicate the request for resources to the at least one of the second edge device or the server to elicit a response from the at least one of the second edge device or the server including resources required to update the trained inference model to enable the updated trained inference model to make an accurate prediction from the captured data;
receive the requested resources;
update the trained inference model using the received resources to enable the updated trained inference model to make an accurate prediction from the captured data; and
make a prediction for the received captured data using the updated, trained inference model.

### Claim 14 (depends on 13)

The apparatus of claim 13, wherein the apparatus is further configured to: identify, in the request for resources, what resources are required to enable the trained inference model to make an accurate prediction from the captured data;
wherein the required resources are determined using a second machine learning model of the first edge device.

### Claim 15 (depends on 13)

The apparatus of claim 13, wherein the request for resources enables the at least one of the second edge device or the server to determine what resources are required to enable the trained inference model to make an accurate prediction from the captured data, using a second learning model.

### Claim 16 (depends on 15)

The apparatus of claim 15, wherein the request for resources comprises at least information regarding the trained inference model and information regarding the captured data.

### Claim 17 (depends on 13)

The apparatus of claim 13, wherein the apparatus is further configured to:
adjust capture parameters of at least one of the at least one sensor to capture additional resources for updating the trained inference model.

### Claim 18 (depends on 13)

The apparatus of claim 13, wherein the server retrieves the required resources from a database comprising heterogeneous models including at least one large model and at least one smaller model, and wherein the database communicates resources to the server based on an available bandwidth between the database and the server.

### Claim 19 (depends on 13)

The apparatus of claim 13, wherein the first edge device retrieves the required resources from a network of heterogenous edge devices including at least one large model and at least one smaller model, and wherein at least one edge device of the network of edge devices communicates resources to the first edge device based on resource constraints of at least one of the first edge device or the edge devices of the network.

### Claim 20 (depends on 13)

The apparatus of claim 13, wherein the apparatus is further configured to:
provide reinforcement learning by monitoring a prediction of the trained inference model after an update and providing a reward to at least one of the first edge device, the second edge device or the server based on a result of the prediction of the trained inference model.

### Claim 21 (depends on 13)

The apparatus of claim 13, wherein the request for the required resources is communicated to the server based on a bandwidth available between the first edge device and at least one of the second edge device or the server at the time of the request.

### Claim 22 (depends on 21)

The apparatus of claim 21, wherein the apparatus is further configured to:
provide reinforcement learning by monitoring communications between the first edge device and at least one of the second edge device or the server and providing a respective reward to at least one of the first edge device, the second edge device or the server based on an amount of bandwidth used for at least each communication request for the required resources.

### Claim 23 (depends on 13)

The apparatus of claim 13, wherein the resources required to update the trained inference model to enable the updated trained inference model to make an accurate prediction from the captured data comprise at least one of data, weights, labels, adaptors, pretrained models, or model updates.

### Claim 24

A system for efficient machine learning with query-based knowledge assistance on an edge device, comprising:
a server;
a database in communication with the server;
a network of edge devices, including at least two edge devices, wherein each edge device includes at least one sensor in communication and wherein each edge device comprises:
a processor; and
a memory accessible to the processor, the memory having stored therein at least one of programs or instructions executable by the processor to configure at least one first edge device of the network of edge devices to:
receive data captured by at least one sensor in communication with the first edge device;
determine a state of the captured data to determine if the captured data includes data that is out of distribution based on a trained inference model of the first edge device;
if the determined state identifies that an amount of the out of distribution data in the captured data is preventing the trained inference model from making an accurate prediction from the captured data, determine a request for resources to be communicated to at least one of a second edge device of the network of edge devices or the server;
communicate the request for resources to the at least one of the second edge device or the server to elicit a response from the at least one of the second edge device or the server including resources required to update the trained inference model to enable the updated trained inference model to make an accurate prediction from the captured data; receive the requested resources;
update the trained inference model using the received resources to enable the updated trained inference model to make an accurate prediction from the captured data; and
make a prediction for the received captured data using the updated, trained inference model.

### Claim 25 (depends on 24)

The system of claim 24, wherein the first edge device is further configured to: identify, in the request for resources, what resources are required to enable the trained inference model to make an accurate prediction from the captured data;
wherein the required resources are determined using a second machine learning model of the first edge device.

### Claim 26 (depends on 24)

The system of claim 24, wherein the request for resources enables the at least one of the second edge device or the server to determine what resources are required to enable the trained inference model to make an accurate prediction from the captured data, using a second learning model.

### Claim 27 (depends on 26)

The system of claim 26, wherein the request for resources comprises at least information regarding the trained inference model and information regarding the captured data.

### Claim 28 (depends on 24)

The system of claim 24, wherein the first edge device is further configured to:
adjust capture parameters of at least one of the at least one sensor to capture additional resources for updating the trained inference model.

### Claim 29 (depends on 24)

The system of claim 24, wherein the server retrieves the required resources from the database, which comprises heterogeneous models including at least one large model and at least one smaller model that communicate resources to the server based on an available bandwidth between the database and the server.

### Claim 30 (depends on 24)

The system of claim 24, wherein the network of edge devices comprise heterogenous edge devices including at least one large model and at least one smaller model.

### Claim 31 (depends on 24)

The system of claim 24, wherein the first edge device is further configured to:
provide reinforcement learning by monitoring a prediction of the trained inference model after an update and providing a reward to at least one of the first edge device, the second edge device or the server based on a result of the prediction of the trained inference model.

### Claim 32 (depends on 24)

The system of claim 24, wherein the request for the required resources is communicated to the server based on a bandwidth available between the first edge device and at least one of the second edge device or the server at the time of the request.

### Claim 33 (depends on 32)

The system of claim 32, wherein the first edge device is further configured to:
provide reinforcement learning by monitoring communications between the first edge device and at least one of the second edge device or the server and providing a respective reward to at least one of the first edge device, the second edge device or the server based on an amount of bandwidth used for at least each communication request for the required resources.

### Claim 34 (depends on 24)

The system of claim 24, wherein the resources required to update the trained inference model to enable the updated trained inference model to make an accurate prediction from the captured data comprise at least one of data, weights, labels, adaptors, pretrained models, or model updates.
