Multi-Precision Convolutional Filters for Neural Network Complexity Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing trainable networks face impractical data and filter storage requirements, memory access bandwidth requirements, and computational complexity, with alternative approaches like pruning and quantization failing to sufficiently reduce resource requirements while maintaining detection or recognition performance.
Innovation Solution
Implementing a multi-layer neural network with convolutional filters of multiple precisions, where higher precision basis filters and lower precision residual filters are used, allowing for reduced computational complexity through clustering and quantization, thereby optimizing resource usage without significant degradation in performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If higher precision filters are used in trainable networks, then detection or recognition performance is improved, but data and filter storage requirements become impractical
Solution Approach 1:
The patent divides filters into multiple precision groups (e.g., high precision and low precision filters) and segments the data processing accordingly. Different precision levels are applied to different filter groups, allowing the system to maintain high detection performance for critical filters while reducing storage requirements for less critical ones.
Solution Approach 2:
The patent applies different precision qualities to different parts of the filter system. High precision is applied locally to filters that contribute most to detection performance, while lower precision is applied to other filters, optimizing the balance between performance and storage requirements.
2Measurement precision
If higher precision filters are used in trainable networks, then detection or recognition performance is improved, but memory access bandwidth requirements become impractical
Solution Approach 1:
The patent segments memory access operations according to filter precision groups. Low precision filters access memory at lower bandwidth rates, while high precision filters receive prioritized access only when needed, reducing overall memory bandwidth requirements while maintaining detection performance.
Solution Approach 2:
The patent applies high precision processing only partially to the most critical filters rather than uniformly to all filters. This partial application of high precision maintains sufficient detection performance while significantly reducing memory bandwidth requirements.
3Measurement precision
If higher precision filters are used in trainable networks, then detection or recognition performance is improved, but computational complexity becomes impractical
Solution Approach 1:
The patent segments computational operations into different precision levels corresponding to different filter groups. Low precision computation is applied to most filters to reduce complexity, while high precision computation is applied selectively to critical filters, maintaining detection performance with reduced overall computational complexity.
Solution Approach 2:
The patent applies high computational precision locally only where detection performance requires it, rather than uniformly across all filters. This localized application of high precision reduces overall computational complexity while maintaining sufficient detection or recognition performance.
4Quantity of substance
If pruning of network connections is applied, then resource requirements are reduced, but detection or recognition performance degrades
Solution Approach 1:
The patent changes the precision parameter of filters and data rather than pruning connections. By using mixed precision (combining high and low precision representations), the system reduces resource requirements through parameter optimization while maintaining detection performance through selective high precision application.
5Quantity of substance
If quantization of coefficients is applied, then resource requirements are reduced, but detection or recognition performance degrades
Solution Approach 1:
The patent applies different quantization levels (precision qualities) to different filters and data elements. Critical filters and data receive higher precision to maintain detection performance, while non-critical elements use lower precision to reduce resource requirements, optimizing the overall system.
Solution Approach 2:
The patent optimizes the precision parameter selectively across different parts of the network rather than applying uniform quantization. This parameter optimization allows resource reduction while maintaining detection performance through strategic high precision application.
Data Source
AI summary
Some embodiments perform, in a multi-layer neural network in a computing device, a convolution operation on input feature maps with multiple convolutional filters. The convolutional filters have multiple filter precisions. In other embodiments, electronic design automation (EDA) systems, methods, and computer-readable media are presented for adding such a multi-layer neural network into an integrated circuit (IC) design.