Kernel Compression via Composite Storage for Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks face increased memory footprint and power consumption due to storing kernels that exhibit similar modifications, such as 90-degree rotations, which require larger memory storage and more power to implement.
Innovation Solution
A method and system for compressing kernels by detecting replicated kernels, generating composite kernels with meta data indicating rotational similarities, and storing these composite kernels, which reduces the number of kernels needed and optimizes memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If replicated kernels are stored separately in convolutional neural networks, then the network can process various rotated features, but the memory footprint and power consumption increase
Solution Approach 1:
The patent merges replicated kernels that are rotated versions of each other into a single composite kernel. Instead of storing multiple separate kernel sets for different rotations, the system combines them into one composite kernel structure that can generate all rotated versions on-demand, thereby reducing memory footprint while maintaining the ability to process various rotated features
Solution Approach 2:
The composite kernel is designed to serve multiple functions simultaneously - it can generate kernels for different rotation angles (0°, 90°, 180°, 270°) as needed. This multi-functional design eliminates the need for separate dedicated kernel stores for each rotation, achieving versatility without proportionally increasing memory resources
2Adaptability or versatility
If replicated kernels are stored separately in convolutional neural networks, then the network can process various rotated features, but the power consumption increases
Solution Approach 1:
By merging replicated kernels into a single composite kernel, the system reduces the total number of memory fetches required. Since memory access is a major power consumer in neural network operations, reducing fetches by having one composite kernel instead of multiple separate kernels directly lowers power consumption while maintaining full rotational feature processing capability
Solution Approach 2:
The patent extracts only the essential unique kernel data into the composite kernel, separating out the rotational transformation information. This allows the system to store minimal unique data and generate the rest through transformation operations, reducing the energy required for data retrieval and processing
3Reliability
If multiple memory fetches are performed for replicated kernels, then the network can access all necessary kernel data, but the processing efficiency decreases
Solution Approach 1:
The composite kernel merges all necessary kernel data into a single unified structure, eliminating the need for multiple separate memory fetches. The system can access all required kernel variations from this single composite structure, maintaining complete data availability while significantly improving processing efficiency by reducing memory access operations
Data Source
AI summary
A method of compressing kernels comprising detecting a plurality of replicated kernels. The plurality of replicated kernels comprise kernels. The method also comprises generating a composite kernel from the replicated kernels. The composite kernel comprises kernel data and meta data indicative of the rotations applied to the composite kernel data. The method also comprises storing a composite kernel.