Neural Network Processing Circuit With Stride-Aware Sparse Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning inference calculations require extensive computational resources, limiting their effectiveness in mobile environments and high-speed processing environments due to high computational complexity.
Innovation Solution
A processing circuit utilizing a stride-aware compressed sparse row (SCSR) algorithm to compress output data of each layer in an artificial neural network by extracting non-zero values based on the stride of subsequent layers, reducing the need for searching valid input pairs and minimizing power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep learning inference calculations are performed using conventional methods, then calculation accuracy is maintained, but computational complexity increases and processing speed decreases
Solution Approach 1:
The patent extracts only the non-zero values from the output data of each layer using a stride-aware compressed sparse row algorithm. By removing zero values that do not contribute to the final result, the data volume is significantly reduced, enabling faster processing and reduced computational complexity in subsequent layers while maintaining calculation accuracy.
Solution Approach 2:
The patent performs compression of output data before it is passed to subsequent layers. By pre-compressing the data using the stride-aware algorithm that leverages knowledge of the next layer's stride, the system prepares optimized input data in advance, reducing the computational burden on subsequent layers and improving overall processing speed.
2Quantity of substance
If conventional compression methods are used, then data volume is reduced, but additional circuits and storage are required for searching valid input pairs
Solution Approach 1:
The compression circuit performs compression operations autonomously using the stride-aware compressed sparse row algorithm without requiring external assistance from separate searching circuits or additional storage units. The circuit leverages the stride information of the next layer to directly identify and compress non-zero values, making the system self-sufficient and eliminating the need for complex additional hardware.
Solution Approach 2:
The compression circuit is designed to handle multiple functions within a single unit: it identifies non-zero values, compresses the data, and prepares it for the next layer all in one operation. This multi-functional approach eliminates the need for separate circuits for searching valid input pairs and storing compressed data, reducing overall device complexity while maintaining data volume reduction benefits.
Data Source
AI summary
Provided are a processing device configured to compress the output data of a first layer into second compressed input data by extracting non-zero values from the output data of the first layer by using a stride-aware compressed sparse row (SCSR) algorithm based on stride of a second layer in the artificial neural network and output the second compressed input data as input data of the second layer.