Condensed Neural Network Filters for Resource-Constrained Devices

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Resource-limited computers face challenges in implementing complex machine learning schemes like convolutional neural networks due to high computational resource requirements.

Innovation Solution

A compact neural network system that uses condensed compound filters to generate multiple individual filters for convolutional layers, sharing weights across filters to reduce resource usage and maintain high-quality performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If complex machine learning schemes like convolutional neural networks are implemented, then high-quality image manipulation and processing performance is achieved, but computational resource requirements (processor power, memory space) increase significantly

Engineering Contradiction:

Improvemachine learning performance qualityVSAvoidcomputational resource requirements

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges multiple individual filters into a single condensed compound filter structure. Instead of storing and executing separate filter operations for each convolutional layer, the system combines their functionality into one unified filter that performs multiple filtering operations simultaneously, reducing memory space and computational overhead while maintaining the same image processing quality

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The condensed compound filter is designed to serve multiple functions that would traditionally require separate filters. A single compound filter can generate multiple individual filters through mathematical operations, allowing one filter structure to perform the work of many, thereby reducing the overall complexity and resource requirements of the neural network

Inventive Principle:

Principle #6Universality (Multi-functionality)

2Adaptability or versatility

If resource-limited computers attempt to store and execute complex machine learning schemes, then comprehensive image manipulation capabilities are obtained, but storage capacity and processing efficiency are compromised

Engineering Contradiction:

Improveimage manipulation capabilitiesVSAvoidmemory space and storage capacity

Core Design Contradiction:

Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The patent extracts the essential filtering functionality from multiple complex individual filters and consolidates it into a single condensed compound filter. By taking out only the necessary computational operations and representing them in a compressed format, the system maintains full image manipulation capabilities while using minimal memory space suitable for resource-limited devices

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The system changes the parameter representation of filters from multiple separate filter sets to a single condensed parameter set. The compound filter uses a compact parameter structure that can be mathematically expanded to generate the equivalent of many individual filters, allowing comprehensive image processing capabilities to be achieved with minimal storage capacity

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12136026B2Compact neural networks using condensed filters

Publication Date: 2024.11.05 SNAP INC

AI summary

A compact neural network system can generate multiple individual filters from a compound filter. Each convolutional layer of a convolutional neural network can include a compound filters used to generate individual filters for that layer. The individual filters overlap in the compound filter and can be extracted using a sampling operation. The extracted individual filters can share weights with nearby filters thereby reducing the overall size of the convolutional neural network.