Neural Network Feature Plane Compression for Embedded Memory Limits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for reducing the circuit scale of neural networks in embedded systems, such as run-length coding and quantization, either limit compression ratio or degrade recognition accuracy.
Innovation Solution
Implementing a data processing apparatus with multiple compression units that perform different types of compression processes on feature plane data, allowing for a higher compression ratio while maintaining recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Volume of stationary object
If run-length coding is used to compress feature planes, then the circuit scale of memory is reduced, but the compression ratio is limited
Solution Approach 1:
The patent segments the compression process into multiple independent compression units, each handling different compression methods (e.g., run-length coding, quantization, pooling). This allows selective application of different compression techniques to different feature map data, achieving higher overall compression ratios while managing memory circuit scale effectively.
Solution Approach 2:
The patent dynamically adjusts compression parameters such as bit-width for quantization and pooling window sizes based on the specific characteristics of feature map data. This adaptive parameter adjustment enables optimization of compression ratio without excessively reducing memory circuit scale, resolving the contradiction between compression efficiency and hardware resource usage.
2Volume of stationary object
If quantization with reduced bit-width is used to compress feature planes, then the circuit scale is reduced, but recognition accuracy deteriorates
Solution Approach 1:
The patent applies different compression strategies to different spatial regions or channels of feature maps based on their importance. Critical regions maintain higher precision while less important regions undergo aggressive compression, thus reducing overall circuit scale while preserving recognition accuracy for key features.
Solution Approach 2:
The patent implements dynamic quantization where the bit-width and compression level are adjusted based on the activation distribution and importance of different feature map elements. This dynamic adaptation allows the system to maintain high recognition accuracy when needed while achieving circuit scale reduction through compression in other scenarios.
3Volume of stationary object
If aggressive compression is applied to reduce memory usage, then memory circuit scale is reduced, but recognition accuracy deteriorates
Solution Approach 1:
The patent performs preliminary analysis of feature map characteristics before applying compression, identifying which regions or channels are critical for recognition accuracy. This preliminary assessment allows the system to apply gentle compression to important data and aggressive compression to less critical data, maintaining reliability while reducing memory circuit scale.
Solution Approach 2:
The patent incorporates feedback mechanisms where compression results are evaluated and used to adjust subsequent compression parameters. If compression causes accuracy deterioration in certain regions, the system automatically adjusts compression levels for those regions in future operations, ensuring recognition accuracy is maintained while achieving memory circuit scale reduction.
Data Source
AI summary
There is provided with a data processing apparatus. An acquisition unit acquires feature plane data of a layer included in a neural network. A control unit outputs a first control signal corresponding to the layer for controlling first compression processing and a second control signal corresponding to the layer for controlling second compression processing. A first compression unit performs the first compression processing corresponding to the first control signal on the feature plane data. A second compression unit performs the second compression processing corresponding to the second control signal on the feature plane data after the first compression processing. A type of processing of the second compression processing is different from the first compression processing.