Multi-Layer Neural Network Filter Precision Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing trainable neural networks face impractical data and filter storage requirements, memory access bandwidth requirements, and computational complexity, with previous approaches like pruning and quantization failing to sufficiently reduce resource demands 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, optimizing chip area and power consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If higher precision filters are used to maintain detection or recognition performance, then recognition accuracy is improved, but computational complexity and resource requirements increase
Solution Approach 1:
The patent segments filters into basis filters and residual filters, allowing different precision levels for different components. Basis filters are stored at lower precision while residual filters capture the remaining information at higher precision, enabling computational complexity reduction while maintaining recognition accuracy through selective precision application.
Solution Approach 2:
The patent applies different precision levels to different parts of the filter representation. Specifically, basis filters use lower precision storage while residual filters use higher precision, creating a local quality differentiation that optimizes the balance between computational complexity and recognition performance for each filter component.
2Device complexity
If filter precision is reduced to decrease computational complexity, then resource requirements are reduced, but recognition performance degrades
Solution Approach 1:
The patent creates a composite filter representation by combining basis filters (lower precision) and residual filters (higher precision). This composite structure allows the system to benefit from both low-precision efficiency and high-precision accuracy, achieving reduced computational complexity while maintaining recognition performance through the synergistic combination of different precision levels.
3Measurement precision
If more filters are used to improve detection performance, then recognition accuracy is improved, but data storage requirements increase
Solution Approach 1:
The patent segments each filter into basis and residual components, enabling efficient storage through selective precision. By storing basis filters at lower precision and only maintaining residual filters where necessary, the system reduces overall data storage requirements while preserving detection performance through the segmented filter representation.
4Productivity
If memory access bandwidth is increased to handle more data, then computational performance is improved, but hardware resource requirements increase
Solution Approach 1:
The patent changes the precision parameter of filter storage, transitioning from uniform high precision to a mixed precision system with basis and residual filters. This parameter change reduces the quantity of data requiring memory access while maintaining computational performance, thereby reducing memory access bandwidth requirements without sacrificing productivity.
Data Source
AI summary
Some embodiments perform, in a multi-layer neural network in a computing device, optimization of the multi-layer neural network, for example by making a convolutional change with a first plurality of convolutional filters, or by making a connection change of a first plurality of convolutional filters. 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.