Pseudo-rounding ANN Parameter Conversion for FPGA Efficiency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing solutions for computing Artificial Neural Networks (ANNs) face challenges with high computational load, power consumption, and inefficiency in processing units like CPUs and GPUs, and lack effective integration with Field Programmable Gate Arrays (FPGAs) for large-scale data movement and global solutions.
Innovation Solution
A system that uses complex rounding rules to convert high-precision ANN parameters to lower precision data types, minimizing vector distance and allowing for iterative adjustments to improve computation quality, which can be implemented in processing units including FPGAs, reducing the need for preserving parameter order and enabling efficient computation with fewer operations and memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high-precision parameters are used in ANN computations, then computation accuracy is improved, but computational load and memory requirements increase
Solution Approach 1:
The patent applies parameter changes by systematically converting ANN parameters from high-precision floating-point format to low-precision fixed-point format. This involves changing the data representation parameters (precision, range, scaling factors) while maintaining computational functionality. The method adjusts quantization parameters and rounding rules to optimize the balance between accuracy and computational efficiency.
2Productivity
If standard rounding rules are applied when converting parameters to lower precision, then computational efficiency is improved, but computation accuracy deteriorates
Solution Approach 1:
The patent implements feedback mechanisms by iteratively evaluating the impact of rounding decisions on computation accuracy. The system monitors quantization errors and adjusts rounding rules accordingly, using performance feedback to refine the conversion process. This allows the system to learn optimal rounding strategies that minimize accuracy loss while maintaining computational efficiency.
Solution Approach 2:
The patent applies dynamics by making the rounding behavior adaptive rather than static. Different rounding rules are applied dynamically based on the specific parameter being converted, its importance in the network, and the desired accuracy level. The system can switch between various rounding modes (rounding to nearest, rounding down, rounding up, random rounding) depending on the context.
3Productivity
If FPGAs are used for ANN computations, then parallel processing capability is improved, but programming complexity increases
Solution Approach 1:
The patent applies universality by developing a general-purpose parameter conversion framework that works across different FPGA architectures and ANN types. The method provides universal rounding rules and conversion techniques that can be applied to various FPGA platforms without requiring architecture-specific customization, thereby reducing programming complexity while maintaining parallel processing capabilities.
Data Source
AI summary
A system for increasing quality of results of computations of an artificial neural network (ANN) by using complex rounding rules for parameters in the ANN is provided, the system comprising one or more processing units configured to: receive a plurality of first parameters for one or more neurons of ANN, the first parameters being of a first data type; and change the first parameters to second parameters of a second data type to obtain a plurality of the second parameters according to a rule in which a distance between at least one first parameter and corresponding second parameter is greater than a distance between the first parameter and a value of the second data type closest to the at least one first parameter. A distance between a vector of the first parameters and a vector of the second parameters is minimized.