Neural Network Pruning for Embedded Surveillance

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Large artificial neural networks with tens or hundreds of millions of parameters are difficult to deploy on embedded systems with limited computing power, as they are often over-parameterized, making it challenging to remove unnecessary parameters without impacting performance.

Innovation Solution

A method that configures an artificial neural network by selecting a subset of object classes relevant to a surveillance situation, accessing activation frequency values, and removing neurons with frequencies lower than a threshold, allowing for adaptive pruning without retraining, which can be combined with other methods like thresholding or regularization to reduce computational resources.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the artificial neural network is expanded with more parameters to improve accuracy, then the recognition performance is improved, but the deployment feasibility on embedded systems deteriorates due to limited computing power

Engineering Contradiction:

Improverecognition accuracyVSAvoidnetwork parameter size

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts and removes unnecessary neurons from the neural network based on activation frequency analysis. By identifying and eliminating neurons that contribute minimally to the output (low activation frequency), the network is pruned to retain only essential parameters, thereby reducing model size while preserving recognition accuracy for targeted object classes.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies different pruning thresholds and activation frequency criteria to different neurons and layers based on their specific contribution to various object classes. This localized approach allows critical neurons for important classes to be preserved while removing less important ones, optimizing the network structure for specific surveillance scenarios rather than applying uniform pruning.

Inventive Principle:

Principle #3Local quality

2Device complexity

If the artificial neural network is pruned to reduce parameter size, then the deployment feasibility on embedded systems is improved, but the recognition performance deteriorates

Engineering Contradiction:

Improvenetwork parameter sizeVSAvoidrecognition accuracy

Core Design Contradiction:

Device complexityVSMeasurement precision

Solution Approach 1:

The patent performs preliminary analysis of neuron activation frequencies during the training phase before deployment. By pre-identifying and marking neurons with low activation frequencies for removal, the system prepares the optimized network structure in advance, ensuring that only essential neurons are retained for maintaining accuracy while reducing overall parameter size.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent uses activation frequency metrics as feedback to guide the pruning process. By continuously monitoring which neurons are activated during training and inference, the system identifies underutilized neurons for removal while preserving those critical for accurate recognition, thereby maintaining performance while reducing model size.

Inventive Principle:

Principle #23Feedback

3Use of energy by moving object

If conventional pruning methods are applied to remove neurons, then the computational resources are reduced, but the adaptability to different surveillance situations deteriorates

Engineering Contradiction:

Improvecomputational resourcesVSAvoidadaptability to surveillance situations

Core Design Contradiction:

Use of energy by moving objectVSAdaptability or versatility

Solution Approach 1:

The patent implements dynamic pruning by allowing the selection of different object class subsets based on specific surveillance situations. The activation frequency thresholds and pruning criteria can be adjusted dynamically according to the required object classes, enabling the network to adapt its structure to different surveillance scenarios while maintaining computational efficiency.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent creates a universal pruning framework that can be applied to different surveillance situations by selecting different subsets of object classes. The same base network can be pruned for different scenarios (e.g., traffic monitoring, security surveillance, wildlife observation) by adjusting which object classes are prioritized, making the solution versatile and adaptable without requiring separate networks for each application.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS10552737B2Artificial neural network class-based pruning

Publication Date: 2020.02.04 AXIS

AI summary

Methods and apparatus, including computer program products, implementing and using techniques for configuring an artificial neural network to a particular surveillance situation. A number of object classes characteristic for the surveillance situation are selected. The object classes form a subset of the total number of object classes for which the artificial neural network is trained. A database is accessed that includes activation frequency values for the neurons within the artificial neural network. The activation frequency values are a function of the object class. Those neurons having activation frequency values lower than a threshold value for the subset of selected object classes are removed from the artificial neural network.