Neural Network Featuremap Compression Using Quad-Tree Bitmaps
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network processors face inefficiencies in data compression and processing due to large data volumes, leading to slow data input and output operations, which hinder real-time data analysis and processing in electronic systems.
Innovation Solution
A neural network processor that employs a compressor to generate compressed data by using location information and data values of non-zero pixels based on a quad-tree structure, reducing data volume and enhancing processing speed by storing compressed data in memory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If featuremap data is stored in memory without compression, then data integrity is maintained, but data volume is large leading to slow input/output operations
Solution Approach 1:
The patent extracts and stores only the essential non-zero pixel data and its location information from the featuremap, discarding redundant zero-value pixel data. This selective extraction reduces data volume while maintaining the information necessary for neural network operations, thereby improving data input/output speed.
Solution Approach 2:
Instead of storing all pixel data including zeros, the patent inverts the approach by storing only the non-zero pixels and their positions. This inverted storage strategy transforms the problem from managing large volumes of mostly redundant data to managing compact representations of significant data points only.
2Productivity
If all pixel data is processed and stored, then complete information is preserved, but processing time increases
Solution Approach 1:
The patent extracts only the necessary non-zero pixel information and location data, eliminating processing overhead for redundant zero-value pixels. This extraction approach reduces the total number of operations required during neural network computations, thereby increasing processing speed and reducing data processing time.
Solution Approach 2:
The patent applies partial action by processing and storing only the essential portion of the data (non-zero pixels) rather than the complete dataset. This partial processing approach is sufficient for maintaining neural network functionality while significantly reducing the time required for data processing.
3Quantity of substance
If compressed data format is used, then data volume is reduced, but compression complexity increases
Solution Approach 1:
The patent segments the featuremap data into distinct components: non-zero pixel values and their corresponding location information. This segmentation allows for a systematic and structured compression approach where each component is handled separately, simplifying the overall compression process while achieving significant data volume reduction.
Solution Approach 2:
The patent changes the representation parameters of the featuremap data by storing pixel locations as coordinates and pixel values as separate entities rather than storing complete pixel arrays. This parameter transformation enables efficient compression by representing data in a more compact form suitable for neural network operations.
Data Source
AI summary
Provided is a neural network device including at least one processor configured to implement an arithmetic circuit configured to generate third data including a plurality of pixels based on a neural network configured to perform an arithmetic operation on first data and second data, and a compressor configured to generate compressed data by compressing the third data, wherein the compressor is further configured to generate, as the compressed data, bitmap data comprising location information about a non-zero pixel having a non-zero data value among the plurality of pixels based on a quad-tree structure.