ANN Quantization via Semifixed-Point Conversion and Fractional Length Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neural networks (ANNs) face a trade-off between accuracy and computational complexity, with floating-point networks offering high accuracy but high resource usage, while fixed-point networks reduce complexity but compromise on accuracy, making efficient quantization methods necessary.
Innovation Solution
A method and apparatus for quantizing ANNs by obtaining quantized parameters and sample distributions, converting floating-point neural networks into semifixed-point networks, and determining optimal fractional lengths to minimize quantization errors, thereby achieving high accuracy with reduced computational resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If floating-point neural networks are used, then accuracy is improved, but computational complexity and resource usage increase
Solution Approach 1:
The patent applies parameter changes by transitioning from floating-point to fixed-point representation for neural network weights and activations. This involves changing the numerical precision parameters while maintaining the network architecture, thereby reducing computational complexity and resource usage while preserving acceptable accuracy levels through careful quantization strategy selection
Solution Approach 2:
The patent segments the quantization process into multiple stages: initial floating-point training, intermediate quantization analysis, and final fixed-point deployment. This segmentation allows the system to identify critical layers and parameters that require higher precision while applying lower precision to less sensitive components, thus resolving the accuracy-complexity trade-off
2Device complexity
If fixed-point neural networks are used, then computational complexity is reduced, but accuracy deteriorates
Solution Approach 1:
The patent implements local quality by applying different quantization precision levels to different layers and parameters within the neural network. Sensitive layers such as early convolutional layers receive higher precision allocation, while less sensitive fully connected layers use lower precision, thereby maintaining overall accuracy while reducing total computational complexity
Solution Approach 2:
The patent performs preliminary quantization analysis during the training phase by monitoring activation distributions and weight sensitivities. This preliminary action enables the system to pre-determine optimal fixed-point parameters before deployment, ensuring that accuracy is preserved while computational complexity is reduced in the final fixed-point model
Data Source
AI summary
An artificial neural network (ANN) quantization method for generating an output ANN by quantizing an input ANN includes: obtaining second parameters by quantizing first parameters of the input ANN; obtaining a sample distribution from an intermediate ANN in which the obtained second parameters have been applied to the input ANN; and obtaining a fractional length for the sample distribution by quantizing the obtained sample distribution.