Hierarchical Hybrid Batch-Incremental Learning for LLNs

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Low power and Lossy Networks (LLNs) face challenges such as lossy links, low bandwidth, limited memory and processing capabilities, and the need for adaptive learning algorithms that can handle changing conditions, where existing batch learning algorithms are computation-intensive and not suited for real-time systems, while incremental algorithms are more adaptable but difficult to tune and prone to local minima.

Innovation Solution

A hierarchical hybrid batch-incremental learning approach where a batch version of the machine learning model runs in a centralized controller and an incremental version is pushed to edge devices, allowing both to operate in parallel, with mechanisms for detecting perturbations and adjusting the forgetting rate to ensure reliable predictions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If batch learning algorithms are used at centralized controllers, then prediction accuracy is improved, but computational complexity and memory requirements increase significantly

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments the learning algorithm into two versions: a batch version running at centralized controllers for high accuracy predictions, and an incremental version running at edge devices for lightweight operation. This segmentation allows each component to use the appropriate complexity level for its specific role and capabilities.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system introduces a hierarchical dimension to the learning architecture, with batch learning operating at the centralized controller level and incremental learning operating at the edge device level. This dimensional separation allows simultaneous use of both high-accuracy and low-complexity approaches in appropriate contexts.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If batch learning algorithms are used, then prediction accuracy is improved, but real-time adaptability deteriorates due to computation intensity

Engineering Contradiction:
Improveprediction accuracyVSAvoidreal-time adaptability
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The learning system is segmented into batch processing at centralized controllers for periodic model updates and incremental processing at edge devices for real-time adaptation. This segmentation enables both high accuracy periodic retraining and fast real-time responsiveness.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The batch learning algorithm performs preliminary action by pre-training models at centralized controllers with complete datasets, creating optimized models that are then deployed to edge devices. This preliminary training enables edge devices to make accurate predictions without performing computationally intensive batch learning in real-time.

Inventive Principle:
Principle #10Preliminary action

3Productivity

If incremental learning algorithms are deployed at edge devices, then computational efficiency is improved, but reliability and convergence performance worsen

Engineering Contradiction:
Improvecomputational efficiencyVSAvoidconvergence performance
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system implements feedback mechanisms where incremental learning results at edge devices are monitored and compared against batch learning baselines. When performance degradation is detected, the system triggers retraining or model updates, ensuring reliability while maintaining computational efficiency.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The batch learning algorithm performs preliminary training to create well-converged models that are then deployed to edge devices. This preliminary convergence action ensures that incremental learning starts from a reliable baseline, improving overall convergence performance while maintaining edge device efficiency.

Inventive Principle:
Principle #10Preliminary action

4Measurement precision

If batch learning is used at centralized controllers, then model accuracy is improved, but memory consumption increases

Engineering Contradiction:
Improvemodel accuracyVSAvoidmemory consumption
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The system segments memory requirements by storing complete datasets and complex models only at centralized controllers with sufficient memory, while edge devices store only lightweight incremental model versions. This segmentation allows high-accuracy batch learning at controllers without burdening edge device memory resources.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3143564B1Hierarchical hybrid batch-incremental learning
Publication Date: 2022.05.04 CISCO TECHNOLOGY INC
  • EP3143564B1 patent drawingFigure 1
  • EP3143564B1 patent drawingFigure 2
  • EP3143564B1 patent drawingFigure 3

AI summary

In one embodiment, a machine learning model for predicting one or more metrics is run in a network which includes a centralized controller device interconnected with a plurality of edge devices. A batch version of the machine learning model that operates in batch mode is hosted at the centralized controller device. Then, an incremental version of the machine learning model that operates in incremental mode is pushed to an edge device of the plurality of edge devices, such that the incremental version of the machine learning model is hosted at the edge device. As a result, the batch version and the incremental version of the machine learning model run in parallel with one another.