Edge Device Dynamic Machine Learning Model Update

Solution Overview

Problem

Static machine learning models at edge devices in distributed systems become stale and fail to effectively learn from local inputs, leading to potential drift away from the hub model and varying performance standards across the system.

Innovation Solution

An edge device computes an objective function based on the relationship between the edge and hub machine learning models, updating the edge model to incorporate knowledge from the hub, thereby mitigating drift and enhancing learning from local data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If static machine learning models are distributed at edge devices, then fast outputs and low-latency system feedback are achieved, but the models become stale and fail to effectively learn from local inputs

Engineering Contradiction:

Improveoutput speedVSAvoidlearning capability

Core Design Contradiction:

SpeedVSAdaptability or versatility

Solution Approach 1:

The patent transforms the static machine learning model into a dynamic model that can continuously update its parameters. The edge device computes an objective function based on the relationship between the edge model and hub model, then updates the edge model parameters dynamically to incorporate knowledge from the hub while adapting to local data distributions, resolving the contradiction between fast static inference and adaptive learning.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent implements a feedback mechanism where the edge device computes an objective function that measures the relationship between the edge model and hub model parameters. This objective function provides feedback signals that guide the update of edge model parameters, enabling the model to continuously learn from local inputs while maintaining alignment with the hub model, thus improving learning capability without sacrificing inference speed.

Inventive Principle:

Principle #23Feedback

2Productivity

If edge devices learn from local data independently, then local learning efficiency improves, but drift away from the hub model occurs

Engineering Contradiction:

Improvelearning efficiencyVSAvoidmodel alignment

Core Design Contradiction:

ProductivityVSStability of the object's composition

Solution Approach 1:

The patent implements a feedback mechanism where the edge device computes an objective function that measures the relationship between the edge model and hub model parameters. This objective function provides feedback signals that guide the update of edge model parameters, enabling the model to continuously learn from local inputs while maintaining alignment with the hub model, thus improving learning capability without sacrificing inference speed.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The patent changes the parameters of the edge machine learning model dynamically based on the computed objective function. By adjusting model parameters through gradient descent or similar optimization methods guided by the objective function, the system enables efficient local learning while controlling drift through parameter updates that incorporate hub model knowledge, resolving the contradiction between learning efficiency and model alignment.

Inventive Principle:

Principle #35Parameter changes

3Stability of the object's composition

If the edge model is updated frequently to prevent drift, then model alignment is maintained, but communication costs and computational overhead increase

Engineering Contradiction:

Improvemodel alignmentVSAvoidcommunication cost

Core Design Contradiction:

Stability of the object's compositionVSLoss of energy

Solution Approach 1:

The patent applies partial action by updating only the necessary parameters of the edge model based on the computed objective function, rather than performing full model synchronization. The edge device computes gradients and updates parameters locally using the objective function as a guide, performing just enough communication and computation to maintain alignment without excessive overhead, thus resolving the contradiction between model alignment and communication cost.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20240273378A1Systems and methods for learning at an edge device

Publication Date: 2024.08.15 ADOBE INC

AI summary

Systems and methods for distributed machine learning are provided. According to one aspect, a method for distributed machine learning includes obtaining, by an edge device, a static machine learning model from a hub device, computing, by the edge device, an objective function for a dynamic machine learning model based on a relationship between the dynamic machine learning model and the static machine learning model, and updating, by the edge device, the dynamic machine learning model based on the objective function.