Feature Map Compression Coding for Neural Network Memory Limits
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning networks face computational and memory intensity due to the complexity of feature maps generated between layers, which reside in memory.
Innovation Solution
A coding apparatus and method that involves lossy and lossless compression of feature maps using K-lossy compression, sparsity-aware quantization, zero-value compression, and block compression to reduce memory burden while maintaining performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If feature maps are stored in memory for deep learning network operations, then computational performance is maintained, but memory requirements become excessively large
Solution Approach 1:
The patent extracts and removes redundant information from feature maps through lossy compression techniques. By identifying and eliminating unnecessary data elements that do not contribute significantly to computational outcomes, the system reduces memory requirements while preserving essential feature information needed for maintaining computational performance.
Solution Approach 2:
The patent applies parameter changes by transforming feature map data through quantization and compression operations. By changing the representation parameters of feature maps (e.g., reducing precision, applying compression algorithms), the system achieves smaller memory footprint while attempting to preserve the critical information needed for computational tasks.
2Quantity of substance
If lossy compression is applied to feature maps, then memory requirements are reduced, but information accuracy deteriorates
Solution Approach 1:
The patent applies partial compression action by selectively compressing only certain portions or aspects of feature maps that are less critical to computational accuracy. By applying compression selectively rather than uniformly, the system reduces overall memory requirements while minimizing information loss in critical regions.
Solution Approach 2:
The patent uses parameter changes through adaptive quantization and compression techniques that adjust the level of compression based on the importance of different feature map regions. By dynamically changing compression parameters, the system balances memory reduction with information preservation.
3Quantity of substance
If compression algorithms are applied to feature maps, then storage efficiency is improved, but computational complexity increases
Solution Approach 1:
The patent segments the feature map processing into distinct compression and decompression stages. By dividing the overall process into separate functional blocks, the system can optimize each stage independently and manage computational complexity more effectively while achieving storage efficiency gains.
Solution Approach 2:
The patent uses copying techniques where compressed feature map representations are created and stored, and then decompressed when needed. This approach allows the system to maintain efficient storage while recovering the necessary information for computation, managing the trade-off between storage efficiency and computational complexity.
Data Source
AI summary
A coding apparatus and a coding method are proposed. The coding apparatus includes a memory and a processor. The processor is configured to obtain a feature map, perform lossy compression on the feature map to generate a lossy feature map, perform lossless compression on the lossy feature map to generate a resultant feature map, and store the resultant feature map in the memory.