Ring Tensor CNN Acceleration Without Irregular Compute Flow
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNN) require high computing power for inference, and existing sparsity techniques like weight pruning cause irregular computing flows and inefficiencies, while methods like quaternion networks and CirCNN increase hardware overheads.
Innovation Solution
A deep neural network accelerating method using ring tensors that perform ring tensor setting, convolution, and non-linear activation functions, employing component-wise products and directional non-linear transforms to maintain image quality without increasing input bitwidths.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If weight pruning is applied to reduce computation demands, then computing power requirements are reduced, but computing flow becomes irregular and control overhead increases
Solution Approach 1:
The patent changes the mathematical representation from standard convolution to ring convolution, where weights are transformed into ring elements. This parameter transformation enables sparsity while maintaining regular computing flow, as the ring structure naturally organizes the computation pattern without requiring complex control logic for pruning operations.
Solution Approach 2:
The patent replaces the mechanical pruning process (removing weights) with a mathematical transformation (ring element representation). Instead of physically removing weights and handling irregular patterns, the system uses ring algebra to represent sparse weights in a way that maintains regular computation flow, substituting a complex control problem with a simpler mathematical formulation.
2Manufacturing precision
If channel shuffling is applied to mix information, then image quality is improved, but input bitwidths of multiplications increase and hardware overhead increases
Solution Approach 1:
The patent applies parameter changes by using ring elements with specific algebraic properties. The ring structure allows channel shuffling to be performed through the multiplication operation itself, where the ring elements' special properties enable information mixing without requiring additional bitwidths. The ring multiplication naturally handles the channel mixing that would otherwise require separate shuffling operations and increased precision.
3Manufacturing precision
If additional linear transform is applied between channels, then image quality improves, but input bitwidths of component-wise products increase
Solution Approach 1:
The patent uses the ring structure to create a mathematical copy of the channel mixing operation. Instead of performing additional linear transforms that would require increased bitwidths, the ring element multiplication provides an equivalent operation that achieves the same information mixing with the original bitwidth preserved. The ring algebra effectively copies the functionality of additional transforms while avoiding the hardware overhead.
Data Source
AI summary
A deep neural network accelerating method using a plurality of ring tensors includes a plurality of steps. A ring tensor setting step includes setting an input feature ring tensor and a weight ring tensor of a convolutional network. A ring tensor convolution calculating step includes calculating a plurality of input feature ring elements of the input feature ring tensor and a plurality of weight ring elements of the weight ring tensor according to a ring multiplication calculating step and a ring addition calculating step to generate a plurality of convolution feature ring elements of a convolution feature ring tensor. A non-linear tensor activation function calculating step includes executing a directional non-linear activation function on one of the convolution feature ring elements of the convolution feature ring tensor to generate an output feature ring element.