Dynamic Resource Allocation for Sensor-Based Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing solutions for deploying machine-learning models on edge devices suffer from accuracy reductions due to model compression techniques, which typically result in performance trade-offs with accuracy losses ranging from 2% to 20%.
Innovation Solution
A system and method for dynamic resource allocation in sensor-based neural networks that share confidence intervals (CIs) across edge devices to optimize processing resources, allowing nodes to focus on tasks based on historical confidence intervals and current resource usage, thereby enhancing decision-making and reducing network latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If model compression techniques are used to deploy machine-learning models on edge devices, then device complexity and processing capability are reduced, but accuracy deteriorates with losses ranging from 2% to 20%
Solution Approach 1:
The system segments the machine learning workload across multiple edge devices rather than relying on a single compressed model. Each device maintains specialized model components or handles specific inference tasks, allowing the system to achieve high accuracy collectively while keeping individual device complexity low.
Solution Approach 2:
Multiple edge devices are merged into a collaborative inference system where their combined processing power compensates for individual model compression. The system aggregates results from multiple devices to achieve overall accuracy that exceeds what any single compressed model could provide.
2Measurement precision
If confidence intervals are shared across edge devices for dynamic resource allocation, then decision-making accuracy is improved, but communication overhead and network latency increase
Solution Approach 1:
The system extracts and shares only the critical confidence interval metadata from each device's local model outputs, rather than exchanging complete model predictions or raw sensor data. This selective extraction minimizes communication overhead while providing sufficient information for accurate collective decision-making.
Solution Approach 2:
Each edge device pre-computes and maintains confidence interval data for its local model predictions before receiving query requests. This preliminary preparation allows the system to quickly aggregate and evaluate confidence intervals from multiple devices without performing computationally intensive calculations during real-time inference, thereby reducing network latency.
3Reliability
If nodes are selected based on greatest confidence interval for processing classification datasets, then reliability of classification is improved, but processing time increases due to evaluating multiple nodes
Solution Approach 1:
The node selection mechanism is dynamic rather than static. The system continuously evaluates confidence intervals from multiple nodes and adaptively selects the most reliable node for each specific classification task based on current confidence metrics. This dynamic selection ensures high reliability while avoiding the overhead of evaluating all nodes for every query by using confidence-based filtering.
Data Source
AI summary
A method, computer program, and computer system are provided for resource allocation for sensor-based neural networks. One or more nodes associated with an edge computing environment are identified. Data corresponding to a classification dataset is received from the identified nodes. The dataset includes a reference classification and confidence value data. A node is selected from among the identified nodes based on the selected node having a greatest confidence interval associated with the reference classification within the confidence value data. The selected node is assigned to process the classification dataset.