Heat-Map-Guided Partial Neural Network Activation on Edge Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
State-of-the-art neural networks require specialized hardware for training and prediction due to their high computational demands, making it impractical to run them on devices with limited processing speed and memory, such as IoT devices and laptops without dedicated accelerators, and they cannot provide real-time deep learning capabilities when disconnected from the cloud.
Innovation Solution
Implementing a partially-activated neural network that activates only a subset of neurons, synapses, and filters based on historical activity patterns encoded in a heat map, optimizing activation frequencies to match historical usage, thereby reducing computational resources and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all neurons, synapses, and filters are activated in each pass (fully-activated neural network), then accuracy and completeness of neural network processing is maintained, but processing speed and computational efficiency deteriorate
Solution Approach 1:
The patent applies partial action by activating only a subset of neurons, synapses, and filters in each pass rather than the entire network. The heat map guides selective activation of only those components that have historically been most active and contribute most to accurate processing, thereby achieving both speed improvement and accuracy maintenance
Solution Approach 2:
The patent implements local quality by differentiating the activation status of different network components based on their historical activity patterns. Components with higher historical activity are activated more frequently, while less active components are activated less frequently, creating a non-uniform activation distribution that optimizes both speed and accuracy
2Productivity
If specialized hardware (GPU) is used for training and runtime, then computational power and processing speed are improved, but device complexity and hardware requirements worsen
Solution Approach 1:
The patent segments the neural network into individually activatable components (neurons, synapses, filters) that can be selectively activated based on heat map guidance. This segmentation enables the network to be executed on general-purpose hardware with limited resources by activating only the necessary segments rather than requiring full specialized hardware capability
Solution Approach 2:
The patent introduces dynamic activation control where the activation pattern changes based on historical activity data. The system adapts which components are activated in each pass, enabling efficient execution on resource-constrained devices by dynamically adjusting the computational footprint rather than requiring static high-power hardware
3Productivity
If continuous communication with cloud is maintained, then access to remote computing power is ensured, but loss of time and connectivity requirements worsen
Solution Approach 1:
The patent enables self-service by allowing the neural network to execute locally on endpoint devices using stored heat map data to guide partial activation. The system serves itself by utilizing locally stored historical activity information to make intelligent activation decisions without requiring continuous cloud connectivity, thereby eliminating latency and connectivity dependencies
4Speed
If high frequency trading or real-time prediction is performed, then processing speed is improved, but use of energy and computational resources worsen
Solution Approach 1:
The patent applies partial action to reduce energy consumption by activating only the necessary subset of network components for each prediction task rather than the entire network. The heat map enables efficient selection of active components, achieving fast real-time predictions with significantly reduced power consumption compared to full network activation
Data Source
AI summary
A device, system, and method for training or prediction of a neural network. A current value may be stored for each of a plurality of synapses or filters in the neural network. A historical metric of activity may be independently determined for each individual or group of the synapses or filters during one or more past iterations. A plurality of partial activations of the neural network may be iteratively executed. Each partial-activation iteration may activate a subset of the plurality of synapses or filters in the neural network. Each individual or group of synapses or filters may be activated in a portion of a total number of iterations proportional to the historical metric of activity independently determined for that individual or group of synapses or filters. Training or prediction of the neural network may be performed based on the plurality of partial activations of the neural network.